Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Personalized Product Recommendation Model of Automatic Question Answering Robot Based on Deep Learning

Personalized Product Recommendation Model of Automatic Question Answering Robot Based on Deep... Hindawi Journal of Robotics Volume 2022, Article ID 1256083, 9 pages https://doi.org/10.1155/2022/1256083 Research Article Personalized Product Recommendation Model of Automatic Question Answering Robot Based on Deep Learning 1 2 Jie Peng and Jianhui Xu Center for Faculty Development, Sichuan Engineering Technical College, Deyang, Sichuan 618000, China Department of Economics and Management, Sichuan Engineering Technical College, Deyang, Sichuan 618000, China Correspondence should be addressed to Jie Peng; pj@scetc.edu.cn Received 30 December 2021; Revised 16 February 2022; Accepted 8 March 2022; Published 21 March 2022 Academic Editor: Shan Zhong Copyright © 2022 Jie Peng and Jianhui Xu. ,is is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. ,e collaborative filtering algorithm widely used in recommendation systems has problems with the sparsity of scoring data and the cold start of new products. A personalized product recommendation model for automated question-answering robots using deep learning is proposed. First, a personalized attention mechanism at the word level and the comment level is proposed, and the comments and users are individually coded. ,en, the bidirectional gated recurrent unit (Bi-GRU) is used to construct the score prediction matrix, and through the dynamic collaborative filtering algorithm to integrate the time characteristics of the user’s interest changes. Finally, the feature codes of the users and products are input into the Bi-GRU model for learning, so as to output the recommendation list of personalized products of the automated question answering robot. Experimental results based on the JD and Tianchi datasets show that the training loss of the proposed model is lower than 45 and 23, respectively. And HR@15 and MRR@15 exceed 48 and 15, respectively, which are better than other comparison models. It can better adapt to the actual needs of automatic question-answering robots. information and according to the score prediction results for 1. Introduction personalized display. ,e deep popularization of the Internet and the rapid de- ,e recommendation algorithm is the core of person- velopment of communication technology have enriched alized recommendation, which directly determines the online service businesses increasingly. ,e subsequent data recommendation performance of the recommendation scale also increased sharply, and the information overload system, and has become a hot issue in current research [4]. was serious. Whether for physical goods or virtual services, Reference [5] explored the consumption habits of hyper- the recommendation system is an important technical personalized health products as unconventional luxury means to solve the problem of service information data goods. Research shows that consumers think hyper-per- overload [1]. Automatic question-answering technology is a sonalized products are worth their money, whether they want to own them or not. ,erefore, potential consumption new intelligent retrieval system that allows users to take natural language query as input, and the system can find and habits are one of the leading factors in the shopping process. return the exact answer from the relevant documents. Es- Although the traditional recommendation algorithm has pecially for self-service platforms, such as automatic ques- developed to a certain extent, it still has certain short- tion-answering robots, the quality of the recommendation comings. Among them, collaborative filtering is a technol- algorithm is directly related to the long-term benefits of its ogy commonly used in personalized recommendation own development [2, 3]. Personalized recommendation is systems. ,e basic idea is to use users with the same interests based on the needs of the users, mining products, or services in the past to choose similar products in the future [6]. that users are interested in from a large amount of However, the algorithm relies too much on historical data, 2 Journal of Robotics communicate with people. By adopting a deep learning has a cold start problem, and performs poorly in the face of new or unpopular products. A typical collaborative filtering method based on recurrent neural network encoder, con- volutional neural network encoder, and bidirectional at- algorithm is a collaborative filtering recommendation model based on matrix factorization. Although this model has good tention flow, better human–computer interaction recommendation performance, the sparsity of scoring data performance is achieved. However, it has not been able to has always restricted the bottleneck of traditional collabo- improve the quality of personalized product recommen- rative filtering [7]. Reference [8] proposed a novel graph- dation services for the time being. Reference [15] proposes a based ranking-oriented recommendation algorithm based personalized ranking of neural graphs. By incorporating the on the influence of paired preferences on recommendation user–item interaction diagram into the embedded learning and directly using the user–item interaction information in diversity. It uses the users’ explicit and implicit feedback to improve the resource allocation model, and matches the the embedded learning, more complex structures can be used in interaction modeling. ,e experimental results prove target users with users with similar preferences to achieve personalized recommendations. However, due to the the effectiveness of the proposed method. However, the applicability of the platform is poor, and it cannot be structure of the model itself, the cold start problem still exists. In addition, because content-based recommendation generally applied to the robot personalized recommendation only relies on the single content information of the product, service platform for automatic question answering. the accuracy of recommending mature products is not high. In summary, in view of the data sparsity and cold start In general, the performance is much lower than the col- problems of traditional recommendation models, a per- laborative filtering algorithm. It is usually used as a sup- sonalized recommendation model based on deep learning is plement to the collaborative filtering algorithm when the proposed. It is applied to the automatic question-answering robot to generate an accurate personalized product rec- data are sparse. Reference [9] proposed a new method of personalized recommendation using rule-based semantic ommendation list. Most recommendation algorithms based on reviews ignore the personalized information of users or reasoning. It can easily and quickly generate practical so- lutions to personalized recommendations. It establishes a products. ,e proposed model uses a personalized attention mechanism to individually encode the users (commodities) connection between the customer and the store by building a recommendation system to provide seamless information and comments. It can extract deep hidden features and exchange. Reference [10] introduced the design and effectively reduce the impact of data sparsity. implementation of a personalized product recommendation model based on user interests. ,e “shopping basket anal- 2. Questions Raised ysis” function model with the Apriori algorithm as the core uses the sales data in the transaction database. It can dig out ,e recommendation system can accurately locate user all kinds of interesting connections between the products interests and commodity characteristics. It is a bridge be- purchased by customers and help businesses develop mar- tween the users and commodities, and realizes the perfect keting strategies. ,e shelves can be reasonably arranged to match between information producers and information guide sales and attract more customers. Although the above consumers. By analyzing the behavior data generated by algorithm has achieved certain results in the field of per- users’ online consumption, the recommendation system can sonalized recommendation, data sparseness and poor model users’ interests and recommend appropriate goods handling of heterogeneous data still exist. under unsupervised training, which is more intelligent and With the continuous development of computer tech- personalized. ,e recommendation system is more and nology, deep learning has become increasingly mature and more applied in network services. After querying and gradually applied to personalized recommendation services. browsing movies on the Douban film platform, the platform Deep learning can obtain useful information from massive will recommend movies that users may be interested in amounts of data to obtain a connection between items and according to users’ tastes. Today’s headline news recom- users. Secondly, it is also possible to pass all kinds of data mendation system can identify users’ interests and hobbies through the same hidden space to ensure the consistent according to the news users usually browse. After users representation of the data [11, 12]. ,erefore, deep learning browse Taobao, Taobao’s recommendation system will push can solve or alleviate the impact of these problems on the “guess you like” products to users. Suppose a user wants to performance of the recommendation system, and improve buy a pair of basketball shoes from Taobao, and enters the efficiency and accuracy of information acquisition. keywords in the search bar, the recommendation system will Reference [13] designed a personalized recommendation recommend some products that the user may be interested system using machine learning, which can recommend that in. students strengthen their leadership and become unique Recommendation algorithm directly affects the perfor- among their peers. ,e model was built using Python flask mance of the recommendation system. A good recom- and Jupyter notebook, and tested using a public dataset and a mendation algorithm can help users quickly locate goals, private dataset. ,e results show that the model has good save a lot of time, and improve user experience. Although the accuracy. However, the algorithm does not work well in recommendation algorithm has been widely used in major actual application scenarios. Reference [14] proposed an websites and software, there is still much room for im- intelligent humanoid robot with self-learning ability based provement. Traditional algorithms have shortcomings. For on deep learning and big data knowledge base to example, the collaborative filtering algorithm relies on Journal of Robotics 3 historical interactive data, has a cold start problem, performs Text processing poorly in the face of new or unpopular products, and the module User recommendation performance decreases significantly when evaluation the interactive data of user products are very sparse. In addition, because content-based recommendation only Bi-GRU depends on the single content information of goods, the accuracy of recommendation for mature goods is not high, Commodity and the performance is far lower than that of the collabo- review rative filtering algorithm in general. Text processing In order to improve the accuracy of the product rec- module ommendation scheme, the proposed model incorporates time series in the collaborative filtering process. ,is reduces Commodity score Dynamic collaborative the impact of the information process in the product rec- forecast filtering ommendation process. At the same time, using the Bi-GRU for data learning can further ensure the effectiveness of the User comment automatic question-answering robot recommendation behavior scheme. Figure 1: Recommended model architecture. 3. Proposed Personalized Commodity Recommendation Model predicting the commodity scoring matrix. Incorporating time series into the collaborative filtering process reduces the 3.1. Overall Framework. ,e key to the current collaborative impact of the information process in the product recom- filtering product recommendation algorithm is the predic- mendation process [19, 20]. tion of product scores, and there are the following three problems: First, there is the problem of sparseness of the scoring matrix. ,e sparsity of review data is a key factor that 3.2. User and Product Hidden Feature Extraction affects the accuracy of the final score prediction of the personalized product recommendation system. ,is is be- 3.2.1. Personalized Comment Encoder. In the user–comment cause the number of products on the platform is far greater network, a comment set R � 􏼈r , r , . . . , r 􏼉 of a user u is u 1 2 D than the number of comments made by a single user. ,e given, where D represents the maximum number of com- comment data present a very obvious sparseness, which in ments in the user–comment set. In particular, each com- turn affects the insufficient amount of relationship between ment r retains only d words. the users and commodities, and between users and users. Using the pre-trained word vector, R is sent to the word ,e accuracy of the prediction results of commodity ratings d×d′ ×k vector mapping layer to obtain R ∈ R , where k is the has been reduced [16]. Second, the cold start problem. ,ere dimension of the word vector. In order to make the word are two main reasons for the cold start of the recommen- information comprehensively consider the context infor- dation system, namely new users and new products. In the mation in the forward and backward directions in the product recommendation platform, new users register every comment, R is sent to the Bi-GRU for encoding, and H � u u day, and new products are also on the shelves. ,e system ′ d×d ×2o (h , h , . . . , h ) ∈ R is obtained. o represents the 1 2 d cannot clarify the interest level of the new products or new output dimension of GRU. Since it is the Bi-GRU here, its users [17, 18]. Finally, there is the issue of information output dimension is 2o. expiration. Over time, a user’s interest preferences or the Since each user (commodity) has a unique identity (ID), popularity of the product will change. Traditional recom- first, use the first Multilayer Perceptron (MLP) to map the ID mendation algorithms do not consider the impact of this into a low-dimensional vector u ∈ R . ,is vector is used to change. capture the personality information of the user’s word level, To solve the above problems, a personalized product and its expression is: recommendation model based on the Bi-GRU and the dynamic collaborative filtering is proposed. Its overall u � ReLU􏼐ω u + b 􏼑, (1) l 1 q 1 structure is shown in Figure 1. where ω represents the weight of the first MLP. b is the bias First, the text information of user reviews and product 1 1 term. u is the ID of user u. reviews is processed by the Bidirectional Encoder Repre- Each user’s word habits and the polarity expressed by sentations from Transformers (BERT) model and the Bi- words have individual characteristics when they post GRU from the transformer to extract the hidden feature comments. In order to make word latent vectors have vectors of the users and commodities, respectively. ,en, the personalized characteristics, it is first necessary to learn a user’s scoring behavior is introduced, and the final product word level attention vector for a certain user u. ,e specific scoring prediction is realized by the TimeSVD++ algorithm. calculation is as follows: ,e model uses the Bi-GRU to capture the hidden feature vectors of the users, and combines the image data to reduce T (2) s � softmax􏼐H P u 􏼑, l u l l the data sparsity and cold start problems in the process of 4 Journal of Robotics T n×1 2o×n T ϑ where u ∈ R is the transposed vector of u . P ∈ R is l l l U � s R . (8) d×d′ ×1 the transition matrix. s ∈ R is the attention score corresponding to d words in d comments. Next, continue to ,e above has introduced the processing process from use s with personalized information to adjust the words of the user–comment set to the user preference vector U in the the comment, and get the implicit expression of d user–comment network. Similarly, in the product review 1×2o comments: network, the product feature vector I ∈ R can also be obtained from the product review collection. R � s ⊗ H , (3) l u d×d ×1 where transpose the last two dimensions of s ∈ R to get 3.3. Commodity Recommendation Model Based on Dynamic T d×1×d s ∈ R . ⊗ represents the batch matrix multiplication, Collaborative Filtering. ,e dynamic collaborative filtering such that after s and H are multiplied, the first dimension d l u algorithm is based on the Singular Value Decomposition remains unchanged, and only the second and third di- (SVD)++ algorithm and adds time series items, and hence mension matrix multiplications are performed. ,erefore, called the TimeSVD++ algorithm. ,e TimeSVD++ algo- the implicit expression of the final d comments is rithm is evolved from a simple factorization model. Assume d×1×2o R ∈ R . In order to facilitate subsequent calculations, N×M that the user’s score prediction matrix y ∈ R for the d×2o the dimension is converted to R ∈ R . product has been obtained. N and M, respectively, represent the number of users and the number of products. y in the ui matrix represents the predicted value of user u’s rating of 3.2.2. Personalized User (Commodity) Encoder. product i. Considering that not all information in R is conducive to ,e deep features of the users and products obtained by constructing a user preference vector, there is a small the Bi-GRU processing review text information are F and amount of irrelevant information. ,erefore, before gath- F , respectively. After coupling with the sharing layer and ering d comments, add a gating mechanism to control the using a factorization machine, the predicted value y of the ui flow of information. Specifically, the input of the gating user’s scoring matrix for the product is obtained. ,e mechanism is R, and its output is a gating weight matrix d×2o TimeSVD++ algorithm first needs to reduce the error be- ϑ ∈ R : tween y 􏽢 and y to obtain the best score prediction. ,e ui ui ϑ � σ Rω + b , (4) specific process can be expressed as: ϑ ϑ 2o×2o y 􏽢 � F F where σ the sigmoid function. ω ∈ R is the weight ui I U matrix. b is the bias term. Next, use ϑ to control the amount 2 (9) min Q � 􏽘 y 􏽢 − y . ui ui of information that each dimension in R can flow into the ui next layer: Adding the bias term in the scoring prediction process (5) R � R ∗ ϑ, constitutes the SVD model: where ∗ is the multiplication of the corresponding ele- y 􏽢 � F F + δ + b + b . (10) ui I U u i ments. Multiply the corresponding elements in R and ϑ to ϑ d×2o obtain the adjusted expression of d comments as R ∈ R . where δ is the average value of the predicted value of the user’s product rating during the training process. b and b In reality, the same expression or similar comments will u i produce different emotional polarities for different users represent the user bias item and the product bias item, respectively, and represent the average value of the score [21, 22]. In order to be able to gather d comments on the prediction value of a user or a product. user’s preference vector based on the user’s personalized Based on the SVD model, the SVD++ algorithm adds information, first, use the second MLP to map the user ID to user interest information through implicit feedback infor- a low-dimensional vector u ∈ R of the review level: mation. In other words, as long as any user has commented u � ReLU􏼐ω u + b 􏼑, (6) r 2 q 2 on a certain product, no matter how high or low the scoring prediction value of the review content is, it means that the where ω is the weight matrix of the second MLP. b is 2 2 user is interested in the product. ,e degree of interest is the bias term. Since different reviews have different con- expressed as c � 􏽮c , c , . . . , c 􏽯 by a hidden factor. At j j1 j2 jQ tributions to the modeling of user preferences [23], it is this time, the user’s rating prediction model for the product necessary to learn the personalized attention vector of the is revised as follows: review level: ϑ T 􏽐 c j∈N(u) j s � softmax􏼐R P u 􏼑, (7) r r 􏽰������ r y 􏽢 � 􏼢F + 􏼣F + δ + b + b , (11) ui U I u i |N(u)| T n×1 2o×n where u ∈ R is the transposed vector of u . P ∈ R is r r r d×1 the transition matrix. s ∈ R is the attention score of each where N(u) represents the set of all products that the user u comment. Next, according to the attention score, the user has evaluated. 1×2o preference vector U ∈ R can be obtained by gathering ,e basic idea of the TimeSVD++ algorithm is that with each comment: the passage of time, the user’s preference for the product Journal of Robotics 5 Table 1: Experimental training parameters. changes, i.e., F , b , and b in equation (11) are no longer U u i fixed quantities, but a function of time. However, the Experimental parameters Specific settings product feature vector F does not change with time. It can Optimization function Adam be seen that the TimeSVD++ algorithm incorporates time Discarding rate 0.6 series items. ,erefore, it is necessary to divide multiple time Learning rate 0.001 periods along the time axis in the process of predicting the Embedded vector dimension 256 user’s rating of the product. Predict product scores in various time periods [24, 25]. ,e divided time period is represented by e(t). ,e values of F , b , and b are the same U u i dataset provides 23291027 interactions of 20,000 customers in the same time period, but are different in different time on 4,758,484 products in one month. periods. ,e scoring prediction process of the TimeSVD++ Before the start of the experiment, the above two datasets algorithm considering time series items is as follows: were preprocessed. First, products with less than 5 ap- pearances and users with less than 10 interactions are 􏽐 c j∈N(u) j 􏽰������ y 􏽢 � 􏼢F (t) + 􏼣F + δ + b (t) + b (t), (12) screened out. ,en, the two datasets are divided into a ui U I u i |N(u)| training set and a test set according to time. 85% of the interactions are the training set, and the rest are the test set, where b (t) and b (t), respectively, represent the bias of the i u i.e., for the JD dataset, 64 days of data are used for training product and the user at time T, and both consist of a static and 11 days of data for testing. For the Tianchi dataset, 26 part and a dynamic part, namely: days of data are used for training, and the remaining 4 days b (t) � b + b i i i,e(t) of data are used for testing. (13) 􏼌 􏼌 β At the same time, in order to control the uniqueness of 􏼌 􏼌 􏼌 􏼌 b (t) � b + β · sign t − s 􏼁 · 􏼌t − s 􏼌 , u u u u u independent variables, when conducting comparative ex- periments, consider that the most commonly used collab- where b represents the offset of the product in the time i,e(t) orative filtering in recommendation cannot recommend period e(t). s represents the average value of all the ratings products that have not appeared before. ,erefore, product given by the user. β represents dynamic weight. interactions and users that have not appeared in the training ,e TimeSVD++ algorithm minimizes the error between set are screened out from the test set. ,e two datasets after the predicted value and the true value after obtaining the being preprocessed are shown in Table 2. predicted value of the score. ,e calculation is as follows: min Q � 􏽘 y 􏽢 − y 􏼁 + ui ui ui 4.2. Evaluation Index. When the recommendation system � � � � � � � � � � � � makes recommendations, a limited number of products will � �2 � �2 � �2 � �2 � � � � � � � � � � � � λ􏼔�F � + �F � + �b (t)� + �b (t)� + c 􏼕, U I u i � j� be recommended each time. If the recommendation is ef- fective, then the products that meet the user’s needs should (14) be included in the recommendation list. Two commonly where λ is the conversion factor. used top-k evaluation indicators are selected to evaluate the quality and effect of the recommendation list generated by the recommendation model. 4. Experiment and Analysis (1) Hit Ratio (HR): HR is calculated as follows: In order to verify the effect of the proposed model, an ex- W@K (15) perimental platform was built using the deep learning HR@K � . framework Tensorflow provided by Google. ,e framework integrates in-depth models such as GRU, which can make ,e meaning of the denominator is all test sets. ,e development simple and easy to understand. ,erefore, it meaning of the numerator is the sum of the number of test has become a more popular learning framework. A three- sets in the top-k list recommended to each user. HR and layer self-attention network will be used in the experiment, Recall have the same functions, and both can evaluate the and other experimental parameter values used are shown in recall rate of the recommendation system in the recall phase. Table 1. It can be seen from the definition that the larger the HR, the better the recommendation effect. Mean Reciprocal Rank (MRR): MRR is calculated as 4.1. Dataset. ,e experiment uses two real datasets, namely follows: the JD dataset under JD and the Tianchi dataset under |N| Alibaba Cloud. ,e JD dataset is provided by the e-com- 1 1 MRR � 􏽘 , (16) merce company Jingdong, one of the largest online B2C N rank i�1 retailers in the country. It contains 370,878,895 interactions with 28,710 products from 105,180 customers in 75 days. ,e where rank is the position of the first item in the ground Tianchi dataset is a public dataset provided by the Ali Mobile truth result of the recommendation list for the ith user. Take recommendation algorithm. It is a real user–commodity MRR@20 as an example. If the item actually clicked by the dataset based on Alibaba’s mobile commerce platform. ,is user appears in the recommendation list and is ranked n th, 6 Journal of Robotics Table 2: Preprocessed dataset parameters. Dataset JD Tianchi Total users 103864 18205 Total commodity 24759 674263 Total interaction 38012329 8665937 Average number of user interactions 451.30 360.94 Training set interaction 32310478 7366046 Test set interaction 5701851 1299891 100 40 70 30 40 20 0 5 10152025 0 5 10152025 Epoch Epoch Ref.[9] Ref.[13] Ref.[9] Ref.[13] Ref.[10] Proposed model Ref.[10] Proposed model (a) (b) Figure 2: Test loss comparison results of different models. (a) JD dataset. (b) Tianchi dataset. recommendation list, which has improved recommendation then n ≥ 1 and if n ≤ 20, then the MRR is equal to 1/n. If this item does not appear in the first 20, the MRR is equal to 0. performance compared to reference [9, 10]. However, there is a lack of dynamic prediction of user interests, and the ,erefore, the larger the MRR, the better. ,e MRR is used to evaluate the order of the items in the recommended list. accuracy of the recommendations needs to be improved. ,erefore, if evaluating the quality of the recommendation sequence of the recommendation model, MRR is a very 4.4. Performance Comparison with Other Algorithms. In important indicator. order to demonstrate the performance of the proposed model, compare it with reference [9, 10, 13]. ,e comparison 4.3. Comparison of Loss Functions. ,e test loss results of the of the evaluation results of selecting HR@15 and MRR@15 proposed model and reference [9, 10, 13] on the twodata sets on the two datasets is shown in Figure 3. In order to see the are shown in Figure 2. ,e ordinate represents the super- change trend of the index more clearly and intuitively, the imposed loss of all data in each epoch. ordinate is enlarged by 100 times. It can be seen from Figure 2 that all four models have On the whole, it can be seen from Figure 3 that the four reached convergence. However, the test loss of the proposed models have great trends in the first 5 iterations. After 5 iterations, the indicator changes stabilized. In the end, the model on the JD and Tianchi datasets is the smallest, which is lower than 45 and 23, respectively. ,e proposed model proposed model’s indicators surpass other models, and the combines the Bi-GRU and the dynamic collaborative fil- effect is the best. Taking HR@15 as an example, the values of tering algorithm to achieve personalized product recom- the proposed models on the two datasets exceed 68 and 48, mendation. ,e hidden attention vector is learned in a respectively. Taking MRR@15 as an example, the values of targeted manner, which can further improve the accuracy of the proposed models on the two datasets exceed 25 and 15, the recommendation. Reference [9] emphasizes rules-based respectively. Reference [9] proposed a rule-based semantic semantic reasoning and reference [10] recommends prod- inference personalized recommendation model, which can ucts based on user interests. All lack dynamic information quickly generate personalized recommendations and prac- filtering, and hence the test loss is relatively large. Reference tical solutions. However, the rules are pre-set and lack [13] uses machine learning to generate a product dynamic updates. ,erefore, the application effect of the Test loss Test loss Journal of Robotics 7 70 30 66 28 62 26 58 24 54 22 50 20 0 5 10152025 0 5 10152025 Epoch Epoch Ref.[9] Ref.[13] Ref.[9] Ref.[13] Ref.[10] Proposed model Ref.[10] Proposed model (a) (b) 55 20 35 10 0 5 10152025 0 510 15 20 25 Epoch Epoch Ref.[9] Ref.[13] Ref.[9] Ref.[13] Ref.[10] Proposed model Ref.[10] Proposed model (c) (d) Figure 3: Comparison results of the evaluation indexes of different models. (a) HR@15 comparison results of the JD dataset. (b) MRR@15 comparison results of the JD dataset. (c) HR@15 comparison results of the Tianchi dataset. (d) MRR@15 comparison results of the Tianchi dataset. automatic question-answering robot is not good. Taking maximum HR@15 should not exceed 67. ,e proposed MRR@15 as an example, which is lower than 25 and 13, model uses the Bi-GRU network to learn text features, and respectively, reference [10] uses user interests to design uses the dynamic collaborative filtering algorithm to gen- personalized product recommendation models. Sales data in erate product recommendations. ,e fused model can better the transaction database is used to mine various interesting adapt to the actual needs of automatic question-answering connections between the products purchased by customers. robots. However, it lacks a powerful learning algorithm as a support, Separately, on the JD dataset, the proposed model has and hence the overall performance is not much different obvious advantages, while on the Tianchi dataset, the from reference [9]. Reference [13] uses machine learning to advantage is slightly smaller. It may be because different build a personalized recommendation system to generate a models of the dataset have different expressiveness be- personalized recommendation plan. Although there are cause the size of the dataset and the data distribution of each latitude in the data are different, resulting in different good learning algorithms for data analysis, there are still some problems in the actual application process because of training effects. ,e proposed model should also be widely the lack of dynamic user interest learning. ,erefore, the used in different datasets to enhance the robustness of the HR@15 HR@15 MRR@15 MRR@15 8 Journal of Robotics model. On the other hand, the average number of in- Conflicts of Interest teractions in the JD dataset is slightly larger than the ,e author declares that there are no conflicts of interest Tianchi dataset. A larger number of interactions can regarding the publication of this paper. capture more information. However, this can also be regarded as a research direction in the future. It is more universal and advantageous to choose and explain the References number of interactions through continuous experimen- [1] P. Nitu, J. Coelho, and P. Madiraju, “Improvising personal- tation. In addition, the existing research experience shows ized travel recommendation system with recency effects,” Big that using data that are closer to the test set time to train Data Mining and Analytics, vol. 4, no. 3, pp. 139–154, 2021. the model will get better results. ,e experimental results [2] D. De Venuto and F. Bellotti, “A CNN based personalized also show that the effect on the JD dataset is better and product recommendation model,” International Journal of more obvious. It also proves the correctness of the pre- Smart Business and Technology, vol. 8, no. 1, pp. 21–32, 2020. vious experience again. [3] J. R. Bock and A. Maewal, “Adversarial learning for product recommendation,” AI, vol. 1, no. 3, pp. 376–388, 2020. 5. Conclusion [4] W. Deng, “Leveraging consumer behaviors for product rec- ommendation: an approach based on heterogeneous net- With the rapid growth of users and product content, au- work,” Electronic Commerce Research, vol. 5, no. 1, pp. 1–27, tomatic question-answering robots using traditional rec- 2020. ommendation algorithms have obvious shortcomings in the [5] M. S. Rosenbaum, G. Contreras Ramirez, J. Campbell, and P. Klaus, “,e product is me: hyper-personalized consumer face of sparse historical interactive data or new product goods as unconventional luxury,” Journal of Business Re- recommendations. ,e use of deep neural networks in the search, vol. 129, no. 6, pp. 446–454, 2021. recommendation field can effectively solve the above [6] B. R. Sreenivasa and C. R. Nirmala, “Hybrid location-centric problems. To this end, a personalized product recommen- e-Commerce recommendation model using dynamic be- dation model for automated question-answering robots havioral traits of customer,” Iran Journal of Computer Science, using deep learning is proposed. A personalized attention vol. 2, no. 3, pp. 179–188, 2019. mechanism is used to personalize comments and users and [7] B. Sheng and G. Sun, “Matrix factorization recommendation we input the Bi-GRU score prediction model for learning. At algorithm based on multiple social relationships,” Mathe- the same time, the dynamic collaborative filtering algorithm matical Problems in Engineering, vol. 2021, Article ID is used to integrate the time characteristics of user interest 6610645, 8 pages, 2021. [8] W. Zhou and W. Han, “Personalized recommendation via changes so as to get the recommendation plan of person- user preference matching,” Information Processing & Man- alized products. ,e experimental results based on the JD agement, vol. 56, no. 3, pp. 955–968, 2019. and Tianchi datasets show that: [9] X. Chen, Y. Xue, and Y. Shiue, “Rule based semantic rea- (1) ,e proposed model uses the attention mechanism soning for personalized recommendation in indoor O2O to obtain in-depth features of products and reviews, e-commerce,” International Core Journal of Engineering, vol. 6, no. 1, pp. 309–318, 2020. which can reduce the error of the recommended [10] J. Zhang, “Personalised product recommendation model solutions. Its training loss is lower than 45 and 23, based on user interest,” Computer Systems Science and En- respectively. gineering, vol. 34, no. 4, pp. 231–236, 2019. (2) ,e proposed model combines the Bi-GRU and the [11] X. Bu, J. Zhu, and X. Qian, “Personalized product search based dynamic collaborative filtering algorithm to achieve on user transaction history and hypergraph learning,” Mul- personalized product recommendation. Compre- timedia Tools and Applications, vol. 79, no. 11, pp. 1–19, 2020. hensive consideration of the timeliness of user in- [12] B. Wu, Y. Ye, and Y. Chen, “Visual appearance or functional complementarity: which aspect affects your decision mak- terest can improve the reliability of the ing?” Information Sciences, vol. 476, no. 8, pp. 19–37, 2019. recommended scheme. Take MRR@15 as an ex- [13] G. Suganya, M. Premlatha, P. Dubey, A. R. Drolia, and ample; its values on the two datasets exceed 25 and S. Srihari, “Subjective areas of improvement: a personalized 15, respectively. ,e overall performance is better recommendation,” Procedia Computer Science, vol. 172, no. 5, than other models. pp. 235–239, 2020. [14] W. Budiharto, V. Andreas, and A. Gunawan, “Deep learning- Experiments have found that due to the complex model based question answering system for intelligent humanoid of the deep learning network, training takes a lot of time and robot,” Journal of Big Data, vol. 7, no. 1, pp. 77–92, 2020. requires high computer configuration. ,erefore, next, we [15] Z. Hu, J. Wang, Y. Yan, P. Zhao, J. Chen, and J. Huang, will consider the optimization and upgradation of the deep “Neural graph personalized ranking for Top-N recommen- neural networks and the combination with the recom- dation,” Knowledge-Based Systems, vol. 213, no. 8, mendation system based on efficiency to ensure that it can be pp. 106426–106437, 2020. more efficient in practical applications. [16] W. Dai, Y. Wang, J. Ma, and Q. Jin, “BTR: a feature-based Bayesian task recommendation scheme for crowdsourcing Data Availability system,” IEEE Transactions on Computational Social Systems, vol. 7, no. 3, pp. 780–789, 2020. ,e data used to support the findings of this study are in- [17] B. Wu, X. He, Z. Sun, L. Chen, and Y. Ye, “ATM: an attentive cluded within the article. translation model for next-item recommendation,” IEEE Journal of Robotics 9 Transactions on Industrial Informatics, vol. 16, no. 3, pp. 1448–1459, 2020. [18] S. Kang, C. Jeong, and K. Chung, “Tree-based real-time ad- vertisement recommendation system in online broadcasting,” IEEE Access, vol. 8, no. 3, pp. 192693–192702, 2020. [19] T. Schreiner, A. Rese, and D. Baier, “Multichannel person- alization: identifying consumer preferences for product rec- ommendations in advertisements across different media channels,” Journal of Retailing and Consumer Services, vol. 48, no. 5, pp. 87–99, 2019. [20] M. Asenova and C. Chrysoulas, “Personalized micro-service recommendation system for online news,” Procedia Computer Science, vol. 160, no. 3, pp. 610–615, 2019. [21] B. Huynh, A. Ibrahim, Y. S. Chang, H. Tobias, and J. O’Donovan, “User perception of situated product recom- mendations in augmented reality,” International Journal of Semantic Computing, vol. 13, no. 3, pp. 289–310, 2019. [22] C. Yu, Z. Zhang, C. Lin, and Y. J. Wu, “Can data-driven precision marketing promote user ad clicks? Evidence from advertising in WeChat moments,” Industrial Marketing Management, vol. 90, no. 5, pp. 481–492, 2020. [23] W. Li and B. Xu, “Aspect-based fashion recommendation with attention mechanism,” IEEE Access, vol. 8, no. 8, pp. 141814–141823, 2020. [24] R. S. Sreepada and B. K. Patra, “Mitigating long tail effect in recommendations using few shot learning technique,” Expert Systems with Applications, vol. 140, no. 2, pp. 112887.1– 112887.17, 2020. [25] S. Hu, A. Kumar, F. Al-Turjman, S. Gupta, S. Seth, and Shubham, “Reviewer credibility and sentiment analysis based user profile modelling for online product recommendation,” IEEE Access, vol. 8, no. 1, pp. 26172–26189, 2020. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Journal of Robotics Hindawi Publishing Corporation

Personalized Product Recommendation Model of Automatic Question Answering Robot Based on Deep Learning

Peng, Jie; Xu, Jianhui
Journal of Robotics , Volume 2022 – Mar 22, 2022
Download PDF
9 pages

Loading next page...
 
/lp/hindawi-publishing-corporation/personalized-product-recommendation-model-of-automatic-question-YVuj0lhxeR

References (26)

Publisher
Hindawi Publishing Corporation
Copyright
Copyright © 2022 Jie Peng and Jianhui Xu. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
ISSN
1687-9600
eISSN
1687-9619
DOI
10.1155/2022/1256083
Publisher site
See Article on Publisher Site

Abstract

Hindawi Journal of Robotics Volume 2022, Article ID 1256083, 9 pages https://doi.org/10.1155/2022/1256083 Research Article Personalized Product Recommendation Model of Automatic Question Answering Robot Based on Deep Learning 1 2 Jie Peng and Jianhui Xu Center for Faculty Development, Sichuan Engineering Technical College, Deyang, Sichuan 618000, China Department of Economics and Management, Sichuan Engineering Technical College, Deyang, Sichuan 618000, China Correspondence should be addressed to Jie Peng; pj@scetc.edu.cn Received 30 December 2021; Revised 16 February 2022; Accepted 8 March 2022; Published 21 March 2022 Academic Editor: Shan Zhong Copyright © 2022 Jie Peng and Jianhui Xu. ,is is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. ,e collaborative filtering algorithm widely used in recommendation systems has problems with the sparsity of scoring data and the cold start of new products. A personalized product recommendation model for automated question-answering robots using deep learning is proposed. First, a personalized attention mechanism at the word level and the comment level is proposed, and the comments and users are individually coded. ,en, the bidirectional gated recurrent unit (Bi-GRU) is used to construct the score prediction matrix, and through the dynamic collaborative filtering algorithm to integrate the time characteristics of the user’s interest changes. Finally, the feature codes of the users and products are input into the Bi-GRU model for learning, so as to output the recommendation list of personalized products of the automated question answering robot. Experimental results based on the JD and Tianchi datasets show that the training loss of the proposed model is lower than 45 and 23, respectively. And HR@15 and MRR@15 exceed 48 and 15, respectively, which are better than other comparison models. It can better adapt to the actual needs of automatic question-answering robots. information and according to the score prediction results for 1. Introduction personalized display. ,e deep popularization of the Internet and the rapid de- ,e recommendation algorithm is the core of person- velopment of communication technology have enriched alized recommendation, which directly determines the online service businesses increasingly. ,e subsequent data recommendation performance of the recommendation scale also increased sharply, and the information overload system, and has become a hot issue in current research [4]. was serious. Whether for physical goods or virtual services, Reference [5] explored the consumption habits of hyper- the recommendation system is an important technical personalized health products as unconventional luxury means to solve the problem of service information data goods. Research shows that consumers think hyper-per- overload [1]. Automatic question-answering technology is a sonalized products are worth their money, whether they want to own them or not. ,erefore, potential consumption new intelligent retrieval system that allows users to take natural language query as input, and the system can find and habits are one of the leading factors in the shopping process. return the exact answer from the relevant documents. Es- Although the traditional recommendation algorithm has pecially for self-service platforms, such as automatic ques- developed to a certain extent, it still has certain short- tion-answering robots, the quality of the recommendation comings. Among them, collaborative filtering is a technol- algorithm is directly related to the long-term benefits of its ogy commonly used in personalized recommendation own development [2, 3]. Personalized recommendation is systems. ,e basic idea is to use users with the same interests based on the needs of the users, mining products, or services in the past to choose similar products in the future [6]. that users are interested in from a large amount of However, the algorithm relies too much on historical data, 2 Journal of Robotics communicate with people. By adopting a deep learning has a cold start problem, and performs poorly in the face of new or unpopular products. A typical collaborative filtering method based on recurrent neural network encoder, con- volutional neural network encoder, and bidirectional at- algorithm is a collaborative filtering recommendation model based on matrix factorization. Although this model has good tention flow, better human–computer interaction recommendation performance, the sparsity of scoring data performance is achieved. However, it has not been able to has always restricted the bottleneck of traditional collabo- improve the quality of personalized product recommen- rative filtering [7]. Reference [8] proposed a novel graph- dation services for the time being. Reference [15] proposes a based ranking-oriented recommendation algorithm based personalized ranking of neural graphs. By incorporating the on the influence of paired preferences on recommendation user–item interaction diagram into the embedded learning and directly using the user–item interaction information in diversity. It uses the users’ explicit and implicit feedback to improve the resource allocation model, and matches the the embedded learning, more complex structures can be used in interaction modeling. ,e experimental results prove target users with users with similar preferences to achieve personalized recommendations. However, due to the the effectiveness of the proposed method. However, the applicability of the platform is poor, and it cannot be structure of the model itself, the cold start problem still exists. In addition, because content-based recommendation generally applied to the robot personalized recommendation only relies on the single content information of the product, service platform for automatic question answering. the accuracy of recommending mature products is not high. In summary, in view of the data sparsity and cold start In general, the performance is much lower than the col- problems of traditional recommendation models, a per- laborative filtering algorithm. It is usually used as a sup- sonalized recommendation model based on deep learning is plement to the collaborative filtering algorithm when the proposed. It is applied to the automatic question-answering robot to generate an accurate personalized product rec- data are sparse. Reference [9] proposed a new method of personalized recommendation using rule-based semantic ommendation list. Most recommendation algorithms based on reviews ignore the personalized information of users or reasoning. It can easily and quickly generate practical so- lutions to personalized recommendations. It establishes a products. ,e proposed model uses a personalized attention mechanism to individually encode the users (commodities) connection between the customer and the store by building a recommendation system to provide seamless information and comments. It can extract deep hidden features and exchange. Reference [10] introduced the design and effectively reduce the impact of data sparsity. implementation of a personalized product recommendation model based on user interests. ,e “shopping basket anal- 2. Questions Raised ysis” function model with the Apriori algorithm as the core uses the sales data in the transaction database. It can dig out ,e recommendation system can accurately locate user all kinds of interesting connections between the products interests and commodity characteristics. It is a bridge be- purchased by customers and help businesses develop mar- tween the users and commodities, and realizes the perfect keting strategies. ,e shelves can be reasonably arranged to match between information producers and information guide sales and attract more customers. Although the above consumers. By analyzing the behavior data generated by algorithm has achieved certain results in the field of per- users’ online consumption, the recommendation system can sonalized recommendation, data sparseness and poor model users’ interests and recommend appropriate goods handling of heterogeneous data still exist. under unsupervised training, which is more intelligent and With the continuous development of computer tech- personalized. ,e recommendation system is more and nology, deep learning has become increasingly mature and more applied in network services. After querying and gradually applied to personalized recommendation services. browsing movies on the Douban film platform, the platform Deep learning can obtain useful information from massive will recommend movies that users may be interested in amounts of data to obtain a connection between items and according to users’ tastes. Today’s headline news recom- users. Secondly, it is also possible to pass all kinds of data mendation system can identify users’ interests and hobbies through the same hidden space to ensure the consistent according to the news users usually browse. After users representation of the data [11, 12]. ,erefore, deep learning browse Taobao, Taobao’s recommendation system will push can solve or alleviate the impact of these problems on the “guess you like” products to users. Suppose a user wants to performance of the recommendation system, and improve buy a pair of basketball shoes from Taobao, and enters the efficiency and accuracy of information acquisition. keywords in the search bar, the recommendation system will Reference [13] designed a personalized recommendation recommend some products that the user may be interested system using machine learning, which can recommend that in. students strengthen their leadership and become unique Recommendation algorithm directly affects the perfor- among their peers. ,e model was built using Python flask mance of the recommendation system. A good recom- and Jupyter notebook, and tested using a public dataset and a mendation algorithm can help users quickly locate goals, private dataset. ,e results show that the model has good save a lot of time, and improve user experience. Although the accuracy. However, the algorithm does not work well in recommendation algorithm has been widely used in major actual application scenarios. Reference [14] proposed an websites and software, there is still much room for im- intelligent humanoid robot with self-learning ability based provement. Traditional algorithms have shortcomings. For on deep learning and big data knowledge base to example, the collaborative filtering algorithm relies on Journal of Robotics 3 historical interactive data, has a cold start problem, performs Text processing poorly in the face of new or unpopular products, and the module User recommendation performance decreases significantly when evaluation the interactive data of user products are very sparse. In addition, because content-based recommendation only Bi-GRU depends on the single content information of goods, the accuracy of recommendation for mature goods is not high, Commodity and the performance is far lower than that of the collabo- review rative filtering algorithm in general. Text processing In order to improve the accuracy of the product rec- module ommendation scheme, the proposed model incorporates time series in the collaborative filtering process. ,is reduces Commodity score Dynamic collaborative the impact of the information process in the product rec- forecast filtering ommendation process. At the same time, using the Bi-GRU for data learning can further ensure the effectiveness of the User comment automatic question-answering robot recommendation behavior scheme. Figure 1: Recommended model architecture. 3. Proposed Personalized Commodity Recommendation Model predicting the commodity scoring matrix. Incorporating time series into the collaborative filtering process reduces the 3.1. Overall Framework. ,e key to the current collaborative impact of the information process in the product recom- filtering product recommendation algorithm is the predic- mendation process [19, 20]. tion of product scores, and there are the following three problems: First, there is the problem of sparseness of the scoring matrix. ,e sparsity of review data is a key factor that 3.2. User and Product Hidden Feature Extraction affects the accuracy of the final score prediction of the personalized product recommendation system. ,is is be- 3.2.1. Personalized Comment Encoder. In the user–comment cause the number of products on the platform is far greater network, a comment set R � 􏼈r , r , . . . , r 􏼉 of a user u is u 1 2 D than the number of comments made by a single user. ,e given, where D represents the maximum number of com- comment data present a very obvious sparseness, which in ments in the user–comment set. In particular, each com- turn affects the insufficient amount of relationship between ment r retains only d words. the users and commodities, and between users and users. Using the pre-trained word vector, R is sent to the word ,e accuracy of the prediction results of commodity ratings d×d′ ×k vector mapping layer to obtain R ∈ R , where k is the has been reduced [16]. Second, the cold start problem. ,ere dimension of the word vector. In order to make the word are two main reasons for the cold start of the recommen- information comprehensively consider the context infor- dation system, namely new users and new products. In the mation in the forward and backward directions in the product recommendation platform, new users register every comment, R is sent to the Bi-GRU for encoding, and H � u u day, and new products are also on the shelves. ,e system ′ d×d ×2o (h , h , . . . , h ) ∈ R is obtained. o represents the 1 2 d cannot clarify the interest level of the new products or new output dimension of GRU. Since it is the Bi-GRU here, its users [17, 18]. Finally, there is the issue of information output dimension is 2o. expiration. Over time, a user’s interest preferences or the Since each user (commodity) has a unique identity (ID), popularity of the product will change. Traditional recom- first, use the first Multilayer Perceptron (MLP) to map the ID mendation algorithms do not consider the impact of this into a low-dimensional vector u ∈ R . ,is vector is used to change. capture the personality information of the user’s word level, To solve the above problems, a personalized product and its expression is: recommendation model based on the Bi-GRU and the dynamic collaborative filtering is proposed. Its overall u � ReLU􏼐ω u + b 􏼑, (1) l 1 q 1 structure is shown in Figure 1. where ω represents the weight of the first MLP. b is the bias First, the text information of user reviews and product 1 1 term. u is the ID of user u. reviews is processed by the Bidirectional Encoder Repre- Each user’s word habits and the polarity expressed by sentations from Transformers (BERT) model and the Bi- words have individual characteristics when they post GRU from the transformer to extract the hidden feature comments. In order to make word latent vectors have vectors of the users and commodities, respectively. ,en, the personalized characteristics, it is first necessary to learn a user’s scoring behavior is introduced, and the final product word level attention vector for a certain user u. ,e specific scoring prediction is realized by the TimeSVD++ algorithm. calculation is as follows: ,e model uses the Bi-GRU to capture the hidden feature vectors of the users, and combines the image data to reduce T (2) s � softmax􏼐H P u 􏼑, l u l l the data sparsity and cold start problems in the process of 4 Journal of Robotics T n×1 2o×n T ϑ where u ∈ R is the transposed vector of u . P ∈ R is l l l U � s R . (8) d×d′ ×1 the transition matrix. s ∈ R is the attention score corresponding to d words in d comments. Next, continue to ,e above has introduced the processing process from use s with personalized information to adjust the words of the user–comment set to the user preference vector U in the the comment, and get the implicit expression of d user–comment network. Similarly, in the product review 1×2o comments: network, the product feature vector I ∈ R can also be obtained from the product review collection. R � s ⊗ H , (3) l u d×d ×1 where transpose the last two dimensions of s ∈ R to get 3.3. Commodity Recommendation Model Based on Dynamic T d×1×d s ∈ R . ⊗ represents the batch matrix multiplication, Collaborative Filtering. ,e dynamic collaborative filtering such that after s and H are multiplied, the first dimension d l u algorithm is based on the Singular Value Decomposition remains unchanged, and only the second and third di- (SVD)++ algorithm and adds time series items, and hence mension matrix multiplications are performed. ,erefore, called the TimeSVD++ algorithm. ,e TimeSVD++ algo- the implicit expression of the final d comments is rithm is evolved from a simple factorization model. Assume d×1×2o R ∈ R . In order to facilitate subsequent calculations, N×M that the user’s score prediction matrix y ∈ R for the d×2o the dimension is converted to R ∈ R . product has been obtained. N and M, respectively, represent the number of users and the number of products. y in the ui matrix represents the predicted value of user u’s rating of 3.2.2. Personalized User (Commodity) Encoder. product i. Considering that not all information in R is conducive to ,e deep features of the users and products obtained by constructing a user preference vector, there is a small the Bi-GRU processing review text information are F and amount of irrelevant information. ,erefore, before gath- F , respectively. After coupling with the sharing layer and ering d comments, add a gating mechanism to control the using a factorization machine, the predicted value y of the ui flow of information. Specifically, the input of the gating user’s scoring matrix for the product is obtained. ,e mechanism is R, and its output is a gating weight matrix d×2o TimeSVD++ algorithm first needs to reduce the error be- ϑ ∈ R : tween y 􏽢 and y to obtain the best score prediction. ,e ui ui ϑ � σ Rω + b , (4) specific process can be expressed as: ϑ ϑ 2o×2o y 􏽢 � F F where σ the sigmoid function. ω ∈ R is the weight ui I U matrix. b is the bias term. Next, use ϑ to control the amount 2 (9) min Q � 􏽘 y 􏽢 − y . ui ui of information that each dimension in R can flow into the ui next layer: Adding the bias term in the scoring prediction process (5) R � R ∗ ϑ, constitutes the SVD model: where ∗ is the multiplication of the corresponding ele- y 􏽢 � F F + δ + b + b . (10) ui I U u i ments. Multiply the corresponding elements in R and ϑ to ϑ d×2o obtain the adjusted expression of d comments as R ∈ R . where δ is the average value of the predicted value of the user’s product rating during the training process. b and b In reality, the same expression or similar comments will u i produce different emotional polarities for different users represent the user bias item and the product bias item, respectively, and represent the average value of the score [21, 22]. In order to be able to gather d comments on the prediction value of a user or a product. user’s preference vector based on the user’s personalized Based on the SVD model, the SVD++ algorithm adds information, first, use the second MLP to map the user ID to user interest information through implicit feedback infor- a low-dimensional vector u ∈ R of the review level: mation. In other words, as long as any user has commented u � ReLU􏼐ω u + b 􏼑, (6) r 2 q 2 on a certain product, no matter how high or low the scoring prediction value of the review content is, it means that the where ω is the weight matrix of the second MLP. b is 2 2 user is interested in the product. ,e degree of interest is the bias term. Since different reviews have different con- expressed as c � 􏽮c , c , . . . , c 􏽯 by a hidden factor. At j j1 j2 jQ tributions to the modeling of user preferences [23], it is this time, the user’s rating prediction model for the product necessary to learn the personalized attention vector of the is revised as follows: review level: ϑ T 􏽐 c j∈N(u) j s � softmax􏼐R P u 􏼑, (7) r r 􏽰������ r y 􏽢 � 􏼢F + 􏼣F + δ + b + b , (11) ui U I u i |N(u)| T n×1 2o×n where u ∈ R is the transposed vector of u . P ∈ R is r r r d×1 the transition matrix. s ∈ R is the attention score of each where N(u) represents the set of all products that the user u comment. Next, according to the attention score, the user has evaluated. 1×2o preference vector U ∈ R can be obtained by gathering ,e basic idea of the TimeSVD++ algorithm is that with each comment: the passage of time, the user’s preference for the product Journal of Robotics 5 Table 1: Experimental training parameters. changes, i.e., F , b , and b in equation (11) are no longer U u i fixed quantities, but a function of time. However, the Experimental parameters Specific settings product feature vector F does not change with time. It can Optimization function Adam be seen that the TimeSVD++ algorithm incorporates time Discarding rate 0.6 series items. ,erefore, it is necessary to divide multiple time Learning rate 0.001 periods along the time axis in the process of predicting the Embedded vector dimension 256 user’s rating of the product. Predict product scores in various time periods [24, 25]. ,e divided time period is represented by e(t). ,e values of F , b , and b are the same U u i dataset provides 23291027 interactions of 20,000 customers in the same time period, but are different in different time on 4,758,484 products in one month. periods. ,e scoring prediction process of the TimeSVD++ Before the start of the experiment, the above two datasets algorithm considering time series items is as follows: were preprocessed. First, products with less than 5 ap- pearances and users with less than 10 interactions are 􏽐 c j∈N(u) j 􏽰������ y 􏽢 � 􏼢F (t) + 􏼣F + δ + b (t) + b (t), (12) screened out. ,en, the two datasets are divided into a ui U I u i |N(u)| training set and a test set according to time. 85% of the interactions are the training set, and the rest are the test set, where b (t) and b (t), respectively, represent the bias of the i u i.e., for the JD dataset, 64 days of data are used for training product and the user at time T, and both consist of a static and 11 days of data for testing. For the Tianchi dataset, 26 part and a dynamic part, namely: days of data are used for training, and the remaining 4 days b (t) � b + b i i i,e(t) of data are used for testing. (13) 􏼌 􏼌 β At the same time, in order to control the uniqueness of 􏼌 􏼌 􏼌 􏼌 b (t) � b + β · sign t − s 􏼁 · 􏼌t − s 􏼌 , u u u u u independent variables, when conducting comparative ex- periments, consider that the most commonly used collab- where b represents the offset of the product in the time i,e(t) orative filtering in recommendation cannot recommend period e(t). s represents the average value of all the ratings products that have not appeared before. ,erefore, product given by the user. β represents dynamic weight. interactions and users that have not appeared in the training ,e TimeSVD++ algorithm minimizes the error between set are screened out from the test set. ,e two datasets after the predicted value and the true value after obtaining the being preprocessed are shown in Table 2. predicted value of the score. ,e calculation is as follows: min Q � 􏽘 y 􏽢 − y 􏼁 + ui ui ui 4.2. Evaluation Index. When the recommendation system � � � � � � � � � � � � makes recommendations, a limited number of products will � �2 � �2 � �2 � �2 � � � � � � � � � � � � λ􏼔�F � + �F � + �b (t)� + �b (t)� + c 􏼕, U I u i � j� be recommended each time. If the recommendation is ef- fective, then the products that meet the user’s needs should (14) be included in the recommendation list. Two commonly where λ is the conversion factor. used top-k evaluation indicators are selected to evaluate the quality and effect of the recommendation list generated by the recommendation model. 4. Experiment and Analysis (1) Hit Ratio (HR): HR is calculated as follows: In order to verify the effect of the proposed model, an ex- W@K (15) perimental platform was built using the deep learning HR@K � . framework Tensorflow provided by Google. ,e framework integrates in-depth models such as GRU, which can make ,e meaning of the denominator is all test sets. ,e development simple and easy to understand. ,erefore, it meaning of the numerator is the sum of the number of test has become a more popular learning framework. A three- sets in the top-k list recommended to each user. HR and layer self-attention network will be used in the experiment, Recall have the same functions, and both can evaluate the and other experimental parameter values used are shown in recall rate of the recommendation system in the recall phase. Table 1. It can be seen from the definition that the larger the HR, the better the recommendation effect. Mean Reciprocal Rank (MRR): MRR is calculated as 4.1. Dataset. ,e experiment uses two real datasets, namely follows: the JD dataset under JD and the Tianchi dataset under |N| Alibaba Cloud. ,e JD dataset is provided by the e-com- 1 1 MRR � 􏽘 , (16) merce company Jingdong, one of the largest online B2C N rank i�1 retailers in the country. It contains 370,878,895 interactions with 28,710 products from 105,180 customers in 75 days. ,e where rank is the position of the first item in the ground Tianchi dataset is a public dataset provided by the Ali Mobile truth result of the recommendation list for the ith user. Take recommendation algorithm. It is a real user–commodity MRR@20 as an example. If the item actually clicked by the dataset based on Alibaba’s mobile commerce platform. ,is user appears in the recommendation list and is ranked n th, 6 Journal of Robotics Table 2: Preprocessed dataset parameters. Dataset JD Tianchi Total users 103864 18205 Total commodity 24759 674263 Total interaction 38012329 8665937 Average number of user interactions 451.30 360.94 Training set interaction 32310478 7366046 Test set interaction 5701851 1299891 100 40 70 30 40 20 0 5 10152025 0 5 10152025 Epoch Epoch Ref.[9] Ref.[13] Ref.[9] Ref.[13] Ref.[10] Proposed model Ref.[10] Proposed model (a) (b) Figure 2: Test loss comparison results of different models. (a) JD dataset. (b) Tianchi dataset. recommendation list, which has improved recommendation then n ≥ 1 and if n ≤ 20, then the MRR is equal to 1/n. If this item does not appear in the first 20, the MRR is equal to 0. performance compared to reference [9, 10]. However, there is a lack of dynamic prediction of user interests, and the ,erefore, the larger the MRR, the better. ,e MRR is used to evaluate the order of the items in the recommended list. accuracy of the recommendations needs to be improved. ,erefore, if evaluating the quality of the recommendation sequence of the recommendation model, MRR is a very 4.4. Performance Comparison with Other Algorithms. In important indicator. order to demonstrate the performance of the proposed model, compare it with reference [9, 10, 13]. ,e comparison 4.3. Comparison of Loss Functions. ,e test loss results of the of the evaluation results of selecting HR@15 and MRR@15 proposed model and reference [9, 10, 13] on the twodata sets on the two datasets is shown in Figure 3. In order to see the are shown in Figure 2. ,e ordinate represents the super- change trend of the index more clearly and intuitively, the imposed loss of all data in each epoch. ordinate is enlarged by 100 times. It can be seen from Figure 2 that all four models have On the whole, it can be seen from Figure 3 that the four reached convergence. However, the test loss of the proposed models have great trends in the first 5 iterations. After 5 iterations, the indicator changes stabilized. In the end, the model on the JD and Tianchi datasets is the smallest, which is lower than 45 and 23, respectively. ,e proposed model proposed model’s indicators surpass other models, and the combines the Bi-GRU and the dynamic collaborative fil- effect is the best. Taking HR@15 as an example, the values of tering algorithm to achieve personalized product recom- the proposed models on the two datasets exceed 68 and 48, mendation. ,e hidden attention vector is learned in a respectively. Taking MRR@15 as an example, the values of targeted manner, which can further improve the accuracy of the proposed models on the two datasets exceed 25 and 15, the recommendation. Reference [9] emphasizes rules-based respectively. Reference [9] proposed a rule-based semantic semantic reasoning and reference [10] recommends prod- inference personalized recommendation model, which can ucts based on user interests. All lack dynamic information quickly generate personalized recommendations and prac- filtering, and hence the test loss is relatively large. Reference tical solutions. However, the rules are pre-set and lack [13] uses machine learning to generate a product dynamic updates. ,erefore, the application effect of the Test loss Test loss Journal of Robotics 7 70 30 66 28 62 26 58 24 54 22 50 20 0 5 10152025 0 5 10152025 Epoch Epoch Ref.[9] Ref.[13] Ref.[9] Ref.[13] Ref.[10] Proposed model Ref.[10] Proposed model (a) (b) 55 20 35 10 0 5 10152025 0 510 15 20 25 Epoch Epoch Ref.[9] Ref.[13] Ref.[9] Ref.[13] Ref.[10] Proposed model Ref.[10] Proposed model (c) (d) Figure 3: Comparison results of the evaluation indexes of different models. (a) HR@15 comparison results of the JD dataset. (b) MRR@15 comparison results of the JD dataset. (c) HR@15 comparison results of the Tianchi dataset. (d) MRR@15 comparison results of the Tianchi dataset. automatic question-answering robot is not good. Taking maximum HR@15 should not exceed 67. ,e proposed MRR@15 as an example, which is lower than 25 and 13, model uses the Bi-GRU network to learn text features, and respectively, reference [10] uses user interests to design uses the dynamic collaborative filtering algorithm to gen- personalized product recommendation models. Sales data in erate product recommendations. ,e fused model can better the transaction database is used to mine various interesting adapt to the actual needs of automatic question-answering connections between the products purchased by customers. robots. However, it lacks a powerful learning algorithm as a support, Separately, on the JD dataset, the proposed model has and hence the overall performance is not much different obvious advantages, while on the Tianchi dataset, the from reference [9]. Reference [13] uses machine learning to advantage is slightly smaller. It may be because different build a personalized recommendation system to generate a models of the dataset have different expressiveness be- personalized recommendation plan. Although there are cause the size of the dataset and the data distribution of each latitude in the data are different, resulting in different good learning algorithms for data analysis, there are still some problems in the actual application process because of training effects. ,e proposed model should also be widely the lack of dynamic user interest learning. ,erefore, the used in different datasets to enhance the robustness of the HR@15 HR@15 MRR@15 MRR@15 8 Journal of Robotics model. On the other hand, the average number of in- Conflicts of Interest teractions in the JD dataset is slightly larger than the ,e author declares that there are no conflicts of interest Tianchi dataset. A larger number of interactions can regarding the publication of this paper. capture more information. However, this can also be regarded as a research direction in the future. It is more universal and advantageous to choose and explain the References number of interactions through continuous experimen- [1] P. Nitu, J. Coelho, and P. Madiraju, “Improvising personal- tation. In addition, the existing research experience shows ized travel recommendation system with recency effects,” Big that using data that are closer to the test set time to train Data Mining and Analytics, vol. 4, no. 3, pp. 139–154, 2021. the model will get better results. ,e experimental results [2] D. De Venuto and F. Bellotti, “A CNN based personalized also show that the effect on the JD dataset is better and product recommendation model,” International Journal of more obvious. It also proves the correctness of the pre- Smart Business and Technology, vol. 8, no. 1, pp. 21–32, 2020. vious experience again. [3] J. R. Bock and A. Maewal, “Adversarial learning for product recommendation,” AI, vol. 1, no. 3, pp. 376–388, 2020. 5. Conclusion [4] W. Deng, “Leveraging consumer behaviors for product rec- ommendation: an approach based on heterogeneous net- With the rapid growth of users and product content, au- work,” Electronic Commerce Research, vol. 5, no. 1, pp. 1–27, tomatic question-answering robots using traditional rec- 2020. ommendation algorithms have obvious shortcomings in the [5] M. S. Rosenbaum, G. Contreras Ramirez, J. Campbell, and P. Klaus, “,e product is me: hyper-personalized consumer face of sparse historical interactive data or new product goods as unconventional luxury,” Journal of Business Re- recommendations. ,e use of deep neural networks in the search, vol. 129, no. 6, pp. 446–454, 2021. recommendation field can effectively solve the above [6] B. R. Sreenivasa and C. R. Nirmala, “Hybrid location-centric problems. To this end, a personalized product recommen- e-Commerce recommendation model using dynamic be- dation model for automated question-answering robots havioral traits of customer,” Iran Journal of Computer Science, using deep learning is proposed. A personalized attention vol. 2, no. 3, pp. 179–188, 2019. mechanism is used to personalize comments and users and [7] B. Sheng and G. Sun, “Matrix factorization recommendation we input the Bi-GRU score prediction model for learning. At algorithm based on multiple social relationships,” Mathe- the same time, the dynamic collaborative filtering algorithm matical Problems in Engineering, vol. 2021, Article ID is used to integrate the time characteristics of user interest 6610645, 8 pages, 2021. [8] W. Zhou and W. Han, “Personalized recommendation via changes so as to get the recommendation plan of person- user preference matching,” Information Processing & Man- alized products. ,e experimental results based on the JD agement, vol. 56, no. 3, pp. 955–968, 2019. and Tianchi datasets show that: [9] X. Chen, Y. Xue, and Y. Shiue, “Rule based semantic rea- (1) ,e proposed model uses the attention mechanism soning for personalized recommendation in indoor O2O to obtain in-depth features of products and reviews, e-commerce,” International Core Journal of Engineering, vol. 6, no. 1, pp. 309–318, 2020. which can reduce the error of the recommended [10] J. Zhang, “Personalised product recommendation model solutions. Its training loss is lower than 45 and 23, based on user interest,” Computer Systems Science and En- respectively. gineering, vol. 34, no. 4, pp. 231–236, 2019. (2) ,e proposed model combines the Bi-GRU and the [11] X. Bu, J. Zhu, and X. Qian, “Personalized product search based dynamic collaborative filtering algorithm to achieve on user transaction history and hypergraph learning,” Mul- personalized product recommendation. Compre- timedia Tools and Applications, vol. 79, no. 11, pp. 1–19, 2020. hensive consideration of the timeliness of user in- [12] B. Wu, Y. Ye, and Y. Chen, “Visual appearance or functional complementarity: which aspect affects your decision mak- terest can improve the reliability of the ing?” Information Sciences, vol. 476, no. 8, pp. 19–37, 2019. recommended scheme. Take MRR@15 as an ex- [13] G. Suganya, M. Premlatha, P. Dubey, A. R. Drolia, and ample; its values on the two datasets exceed 25 and S. Srihari, “Subjective areas of improvement: a personalized 15, respectively. ,e overall performance is better recommendation,” Procedia Computer Science, vol. 172, no. 5, than other models. pp. 235–239, 2020. [14] W. Budiharto, V. Andreas, and A. Gunawan, “Deep learning- Experiments have found that due to the complex model based question answering system for intelligent humanoid of the deep learning network, training takes a lot of time and robot,” Journal of Big Data, vol. 7, no. 1, pp. 77–92, 2020. requires high computer configuration. ,erefore, next, we [15] Z. Hu, J. Wang, Y. Yan, P. Zhao, J. Chen, and J. Huang, will consider the optimization and upgradation of the deep “Neural graph personalized ranking for Top-N recommen- neural networks and the combination with the recom- dation,” Knowledge-Based Systems, vol. 213, no. 8, mendation system based on efficiency to ensure that it can be pp. 106426–106437, 2020. more efficient in practical applications. [16] W. Dai, Y. Wang, J. Ma, and Q. Jin, “BTR: a feature-based Bayesian task recommendation scheme for crowdsourcing Data Availability system,” IEEE Transactions on Computational Social Systems, vol. 7, no. 3, pp. 780–789, 2020. ,e data used to support the findings of this study are in- [17] B. Wu, X. He, Z. Sun, L. Chen, and Y. Ye, “ATM: an attentive cluded within the article. translation model for next-item recommendation,” IEEE Journal of Robotics 9 Transactions on Industrial Informatics, vol. 16, no. 3, pp. 1448–1459, 2020. [18] S. Kang, C. Jeong, and K. Chung, “Tree-based real-time ad- vertisement recommendation system in online broadcasting,” IEEE Access, vol. 8, no. 3, pp. 192693–192702, 2020. [19] T. Schreiner, A. Rese, and D. Baier, “Multichannel person- alization: identifying consumer preferences for product rec- ommendations in advertisements across different media channels,” Journal of Retailing and Consumer Services, vol. 48, no. 5, pp. 87–99, 2019. [20] M. Asenova and C. Chrysoulas, “Personalized micro-service recommendation system for online news,” Procedia Computer Science, vol. 160, no. 3, pp. 610–615, 2019. [21] B. Huynh, A. Ibrahim, Y. S. Chang, H. Tobias, and J. O’Donovan, “User perception of situated product recom- mendations in augmented reality,” International Journal of Semantic Computing, vol. 13, no. 3, pp. 289–310, 2019. [22] C. Yu, Z. Zhang, C. Lin, and Y. J. Wu, “Can data-driven precision marketing promote user ad clicks? Evidence from advertising in WeChat moments,” Industrial Marketing Management, vol. 90, no. 5, pp. 481–492, 2020. [23] W. Li and B. Xu, “Aspect-based fashion recommendation with attention mechanism,” IEEE Access, vol. 8, no. 8, pp. 141814–141823, 2020. [24] R. S. Sreepada and B. K. Patra, “Mitigating long tail effect in recommendations using few shot learning technique,” Expert Systems with Applications, vol. 140, no. 2, pp. 112887.1– 112887.17, 2020. [25] S. Hu, A. Kumar, F. Al-Turjman, S. Gupta, S. Seth, and Shubham, “Reviewer credibility and sentiment analysis based user profile modelling for online product recommendation,” IEEE Access, vol. 8, no. 1, pp. 26172–26189, 2020.

Journal

Journal of Robotics – Hindawi Publishing Corporation

Published: Mar 22, 2022

There are no references for this article.