Multi-pair sorting recommendation algorithm based on multiple implicit feedbacks

Abstract

The invention relates to a multi-pair sorting recommendation algorithm based on various implicit feedbacks, which comprises the following steps of: 1, a data acquisition and processing stage: acquiring user behavior data from a business website and processing the user behavior data; 2, a recommendation algorithm processing stage: training a recommendation algorithm model by using the processed data, and generating a recommendation model; and step 3, a recommendation list stage: predicting scores of the users on the articles by using the recommendation model, sorting the scores of the articles, and finally generating a recommendation list for each user. The multi-pair sorting recommendation algorithm based on multiple implicit feedbacks has the advantages that multiple implicit feedback data, namely purchase behavior, click behavior and shopping cart adding behavior data, are acquired, so that the data volume is enriched to solve the problem of data sparsity; the item popularity method is improved by using various implicit feedback data to mine items which users do not like, and deeper user preferences are provided.

Description

Technical Field

[0001] The present invention relates to a recommendation algorithm for Internet data, and more particularly to a multiple-pair ranking recommendation algorithm based on multiple implicit feedbacks. Background Art

[0002] According to the introduction in the literature 1 (Wodecki A, Wodecki A. Artificial intelligence methods and techniques [J]. Artificial Intelligence in Value Creation: Improving Competitive Advantage, 2019: 71-132.), in recent years, with the rapid development of Internet high-tech, artificial intelligence as its core branch, the development of technologies such as image processing and recommendation systems included in it is changing with each passing day. At the same time, more and more information appears on the Internet, and mankind has entered the era of information explosion.

[0003] According to literature 2 and 3 (Zheng Qiang, Gao Qun. A review of big data research [J]. Science and Technology Vision, 2018, (30): 179-180. 2018.30.078. and literature 3 (Eppler MJ, Mengis J. The Concept of Information Overload-A Review of Literature from Organization Science, Accounting, Marketing, MIS, and Related Disciplines (2004) The Information Society: An International Journal, 20 (5), 2004, pp. 1–20 [J]. Kommunikationsmanagement imWandel: A survey by the Australian Institute of Statistics (AUS) (Jahren=McMinstitute, 2008:271-305) shows that China is one of the countries with the largest amount of data and the most types of data. People can find more information through the Internet, but this also makes it difficult for people to find information that is valuable to them. This phenomenon is often called information overload.

[0004] The author of document 4 (Xiang Liang. Recommendation System Practice [M]. People's Posts and Telecommunications Press, 2012) pointed out that in order to effectively utilize data information, the concept of recommendation system was proposed, which provides users with item or information recommendation services by utilizing collected data; the author of document 5 (Zhang S, Yao L, Sun A, et al. Deep learning based recommender system: A survey and new perspectives [J]. ACM computing surveys (CSUR), 2019, 52 (1): 1-38.) believes that the core of effective utilization of data information is to use the collected data to analyze user characteristics and item characteristics, and then recommend content that users may be interested in. According to the data, the recommendation system can provide better services to users. In the Netflix application, the recommendation system has an 80% impact on users' movie choices, and in YouTube, the recommendation system has a 60% impact on users' video clicks. At the same time, the application scenarios of recommendation systems are very wide. For example, the authors of Reference 6 (Oh J, Kim S, Kim J, et al. When to recommend: A new issue on TV show recommendation [J]. Information Sciences, 2014, 280: 261-274.) and Reference 7 (Steck H, Baltrunas L, Elahi E, et al. Deep learning for recommender systems: A Netflix case study [J]. AIMagazine, 2021, 42 (3): 7-18.) believe that the application scenarios of recommendation systems have been extended to online shopping and music applications. For users, recommendation systems can improve their user experience, and for merchants, recommendation systems can increase their profits, so recommendation systems are very important.

[0005] Currently, according to Document 8 (Deshpande M, Karypis G. Item-based top-n recommendation algorithms [J]. ACM Transactions on Information Systems (TOIS), 2004, 22(1): 143-177.), Document 9 (Koren Y. Factorization meets the neighborhood: a multifacetedcollaborative filtering model [C] / / Proceedings of the 14th ACM SIGKDDinternational conference on Knowledge discovery and data mining.2008:426-434.) and literature 10 (Kunaver M, T.Diversity in recommender systems–A survey[J].Knowledge-based systems,2017,123:154-162.) reported that the main methods included in the recommendation system are content-based recommendation, collaborative filtering recommendation and hybrid recommendation. Among them, collaborative filtering recommendation can be subdivided into memory-based recommendation and model-based recommendation. In recent years, after the evolution of theory and practice, the recommendation system has made great progress. However, according to the report in literature 11 (Wang H, Wang Z, Zhang W.Quantitative analysis of Matthew effect andsparsity problem of recommender systems[C] / / 2018IEEE 3rd InternationalConference on Cloud Computing and Big Data Analysis(ICCCBDA).IEEE,2018:78-82.), there is still a data sparsity problem in the recommendation system, which reduces the performance of the recommendation system and affects the user experience. The data sparsity problem refers to the situation that when the number of users and items is very large, the number of items rated by users is very small, resulting in a very sparse user-item rating matrix, which will greatly reduce the performance of the recommendation system. Therefore, it is very important to overcome the data sparsity problem and improve the recommendation performance.

[0006] In the recommendation system, references 12 (Li Q, Zheng X. Deep collaborative autoencoder for recommender systems: a unified framework for explicit and implicit feedback [J]. arXiv preprint arXiv:1712.09043, 2017.) and 13 ( -Valdez ER, Quintana D, Crespo RG, et al. Are recommender system based on implicit feedback for selective dissemination of ebooks [J]. Information Sciences, 2018, 467: 87-98.) The authors all believe that the most effective information is user feedback, through which we can understand the user's personal preferences. User feedback is divided into two types, one is explicit feedback and the other is implicit feedback. Through explicit feedback, we can clarify the user's preference for the item. The most common explicit feedback in the recommendation system is the user's rating of the item. However, it is impossible to clarify the user's preference for the item through implicit feedback, and we can only determine which items the user may like. Reference 14 (Han J, Zheng L, Huang H, et al. Deep latent factor model with hierarchical similarity measure for recommender systems [J]. Information Sciences, 2019, 503: 521-532.) believes that common implicit feedback in the recommendation system includes user browsing history, likes and some other behaviors.

[0007] In the real world, implicit feedback data is often richer than explicit feedback, such as in Reference 15 (Adomavicius G, Tuzhilin A. Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions [J]. IEEE transactions on knowledge and data engineering, 2005, 17 (6): 734-749.), Reference 16 (Agarwal D, Chen B C. Regression-based latent factor models [C] / / Proceedings of the 15th ACMSIGKDD international conference on Knowledge discovery and data mining. 2009: 19-28.), Reference 17 (Amer-Yahia S, Roy SB, Chawlat A, et al. Group recommendation: Semantics and efficiency [J]. Proceedings of the VLDB Endowment, 2009, 2 (1): 754-765.), Reference 18 (Berry M W. SVDPACK: A Fortran-77software library for the sparsesingular value decomposition[M]. University of Tennessee, 1992.), Document 19 (Bobadilla J, Ortega F, Hernando A, et al. Recommender systems survey [J]. Knowledge-based systems, 2013, 46: 109-132.) and Document 20 (Xu S, Wu S, WangL.The authors of Personalized semantic ranking for collaborative recommendation[C] / / Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval.2015:971-974.) believe that the tasks of recommendation systems can be divided into two types: rating prediction and ranking prediction. In the rating prediction task, explicit rating data is required, but in the real world, such data is often very sparse, and there is a cold start problem, which makes the rating prediction effect not very ideal. On the contrary, the authors of literature 21 (Ding J, Feng F, He X, et al. An improved sampler for bayesian personalized ranking by leveraging view data[C] / / Companion Proceedings of the The Web Conference 2018.2018:13-14.) believe that the ranking prediction task utilizes implicit feedback, and implicit feedback data is easier to obtain and richer than explicit feedback data. Therefore, the ranking prediction method based on implicit feedback is currently a popular research field. However, as a price for availability, implicit feedback information can only reflect the user's unilateral preferences and lacks negative feedback. If the method of introducing implicit feedback is inappropriate or the implicit feedback introduced is too sparse, the recommendation performance will be reduced.

[0008] At present, most recommendation algorithms based on implicit feedback only focus on one type of implicit feedback, such as purchasing behavior in e-commerce. However, in the real world, users generally only buy a few items, which leads to data sparsity. Other types of implicit feedback can reflect user preferences to a certain extent, so it is worthwhile to combine multiple types of implicit feedback.

[0009] Some traditional recommendation algorithms have combined multiple implicit feedbacks, but there are still some problems. In most cases, the target behavior feedback in the recommendation system has data sparsity. For example, in a shopping mall system, users will only buy a small number of items. In fact, in addition to the target behavior feedback that can reflect the user's preference information, the auxiliary behavior feedback can also reflect the user's preference information to a certain extent, and its data volume is also richer. Therefore, combining multiple implicit feedbacks in the recommendation system is one of the important ways to alleviate data sparsity.

[0010] Among the recommendation algorithms based on implicit feedback, the Bayesian algorithm is the most classic recommendation algorithm based on paired learning. The preference assumption proposed in the algorithm still has certain defects. It assumes that items that users have not observed are negative feedback, but users have other possibilities for unobserved items, such as users have not discovered the items or users are not sure about purchasing items. Therefore, by combining multiple implicit feedback to further classify unobserved items, the performance of paired recommendation algorithms can be effectively improved.

[0011] How to combine multiple implicit feedbacks. Currently, many algorithms try to improve the performance of recommendation algorithms by introducing auxiliary behaviors, such as optimizing algorithm performance by combining browsing data as auxiliary behaviors. However, such algorithms often only combine one auxiliary behavior and the preference assumptions have certain defects, which makes the recommendation performance not optimal. Therefore, the reasonable combination of multiple implicit feedbacks can effectively improve the performance of recommendation algorithms.

[0012] In the existing published patent application documents, for example, Chinese invention patent application number CN202410018048.9 discloses a method and system for pushing scientific and technological intelligence information. The method and system for pushing scientific and technological intelligence information described in the invention use a sorting recommendation algorithm to intelligently screen, evaluate and sort the collected information, and put the information that users are more concerned about in the front, thereby helping users to filter out the most popular, most relevant and most timely information from massive news information, and providing users with more efficient, more accurate and more professional energy science and technology intelligence services.

[0013] For another example, Chinese invention patent application No. 201711117106.X discloses an APP recommendation method and system based on APP usage records, the method comprising the following steps: step one, obtaining the user's APP usage records; step two, preprocessing the obtained APP usage records into the required format; step three, generating vector representations of the user and the APP based on the preprocessing results of the APP usage records, and using a joint ranking recommendation algorithm based on the vector representation to recommend K APPs to each user. The present invention can solve the data sparsity problem of fewer APP ratings on smart terminals and improve the accuracy of APP recommendations.

[0014] None of the above invention patent applications solves the sparse problem of recommending information that may be of interest to potential customers who may purchase goods. Summary of the invention

[0015] The purpose of the present invention is to overcome the defects in the prior art and provide a multiple-pair ranking recommendation algorithm based on multiple implicit feedbacks.

[0016] The multiple-pair ranking recommendation algorithm based on multiple implicit feedbacks includes:

[0017] Step 1. Data collection and processing stage: obtaining user behavior data from business websites and processing them;

[0018] Step 2. Recommendation algorithm processing stage: use the processed data to train the recommendation algorithm model and generate a recommendation model;

[0019] Step 3. In the recommendation list stage, the recommendation model is used to predict the user's ratings of items, and the item ratings are sorted, and finally a recommendation list is generated for each user.

[0020] Furthermore, in the data collection preprocessing stage described in step 1, user behavior data is obtained from the business website and processed, specifically including:

[0021] Step 1.1 Obtain the behavior data of cooperative users from the e-commerce website, process the data, filter and clean it to obtain implicit feedback data, which includes purchase data, click data and add to shopping cart data;

[0022] Step 1.2 Generate user-item matrix data based on purchase data, click data, and add-to-cart data, improve the item popularity algorithm based on the algorithm, and use it as a training set; and mine items that users do not like as auxiliary data;

[0023] Step 1.3 In order to make the user's preference assumption more reasonable, the popularity algorithm proposed by Steck H. is introduced to mine items that users do not like. The popularity of an item is defined as the number of users who interact with it. The more users interact with an item, the more popular the item is, and the higher the popularity of the item is. Considering multiple implicit feedbacks, that is, not only considering the impact of target behavior data on item popularity, but also considering the impact of auxiliary behavior data on item popularity, the popularity threshold ε is set p , we get a low-popularity item set I lp , as shown in formula (1):

[0024] I lp ={i∈I - |N i <ε p}......(1),

[0025] Among them, N i represents the number of users who have behavior on item i, when ε p When set to 1, find items with less than 1 user interaction and assign them to I lp For each user, the set of disliked items is obtained by the following formula (2):

[0026]

[0027] Based on auxiliary data, the items are further classified into the target behavior item set Add to cart item set Click Behavior Item Set Disliked Items and the remaining unknown set of items

[0028] Step 1.4 uses implicit feedback data and auxiliary data as training sets:

[0029] Sort the user's preferences for different items, and the items come from purchase data, click data, add to cart data and auxiliary data respectively. Among them, indicates that the user prefers purchased items over unknown items, the user prefers clicked items over unknown items, the user prefers added to cart items over unknown items, the user prefers purchased items over auxiliary items, the user prefers clicked items over auxiliary items, and the user prefers added to cart items over auxiliary items.

[0030] Furthermore, the step 2 uses the user behavior data processed in the data collection and processing stage to train the recommendation algorithm and generate a recommendation model, including:

[0031] Step 2.1: Combine the preference ranking pairs with the matrix decomposition model and the BPR model respectively, and initialize the model parameters. The parameters of the initialized model include: model parameter θ, learning rate η, feature vector dimension d, and regularization coefficient α;

[0032] Step 2.2 uses the sigmoid function to approximate the probability function P(·), and introduces the logarithm to reduce the complexity of the calculation, and finally obtains the objective function of the algorithm;

[0033] Step 2.3 uses stochastic gradient descent to learn the optimal value of the model parameter θ, traverses each preference ranking pair in the data set, and iterates the model and its parameters using the stochastic gradient descent algorithm;

[0034] Through the cyclic gradient descent algorithm, convergence iteration is achieved and the recommendation model is generated, that is, the optimal parameter setting of the proposed model is obtained through training.

[0035] Furthermore, in step 3, in the recommendation list stage, the recommendation model is used to predict and sort the items based on the user behavior data and auxiliary data, specifically including:

[0036] Substitute user behavior data and auxiliary data into the recommendation model to generate a personalized recommendation list for each user: The parameters of the recommendation model are u, i, v, c, j, k, which are expressed as:

[0037] u is the user;

[0038] i is the item purchased by user u;

[0039] v is the item clicked by user u;

[0040] c is the item added to the shopping cart by user u;

[0041] j is an unknown item to user u;

[0042] k is the item that user u dislikes;

[0043] The following 6-tuple formula (3) is used to express the user's behavior preference of liking or disliking to buy items, click item v, and add items to the shopping cart relative to unknown items and disliked items:

[0044]

[0045] In formula (3), the 6-tuple formula indicates that user u prefers to buy item i relative to unknown item j, user u prefers to click item v relative to unknown item j, user u prefers to add item c to the shopping cart relative to unknown item j, user u prefers to buy item i relative to disliked item k, user u prefers to click item v relative to unknown item k, and user u prefers to add item c to the shopping cart relative to unknown item k;

[0046] in: is the set of items for purchase behavior, is the set of items for user click behavior, The set of items that users add to their shopping carts. For the set of disliked items, is the unknown set of items;

[0047] The sigmoid function is used to approximate the probability function P(·), and the logarithm is introduced to reduce the complexity of the calculation. Finally, the objective function of the recommendation algorithm is obtained as follows (5):

[0048]

[0049] Where: R(θ) is a regularization term used to prevent the algorithm from overfitting, as shown in the following formula (6):

[0050] λ u and λ v are the regularization parameters for users and items respectively:

[0051]

[0052] Here, the matrix decomposition framework is adopted, with a total of 6 algorithm parameters, the user feature vector U u and item feature vector V i 、Vv 、V c 、V j and V k , where i, v, c, j and k come from and Set, θ is used to represent the parameters used in MPIF, and the specific parameters are shown in formula (7):

[0053]

[0054] The algorithm parameters are learned by using the stochastic gradient descent method. The update rule for each variable is as follows:

[0055]

[0056] Where η is the learning rate, which is used to control the step size of each iteration and the gradient of each variable, as shown in the following formula (9):

[0057]

[0058] Compared with the prior art in this technical field, the present invention has at least the following beneficial effects:

[0059] 1. The multi-pair ranking recommendation algorithm based on multiple implicit feedbacks described in the present invention obtains multiple implicit feedback data, namely, purchase behavior, click behavior and add-to-cart behavior data, so as to enrich the data volume to solve the data sparsity problem.

[0060] 2. The multi-pair ranking recommendation algorithm based on multiple implicit feedbacks described in the present invention utilizes multiple implicit feedback data to improve the item popularity method to mine items that users dislike and proposes deeper user preferences, that is, it simultaneously considers the relationship between purchase behavior items, click behavior items, add to cart behavior items, disliked items and unobserved items. This method of considering multiple feedback data utilizes the characteristics of large data volume, compensates for data sparsity and preference inaccuracy, and thus improves the accuracy of recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is a schematic diagram of MPIF item classification of the multi-pair ranking recommendation algorithm based on multiple implicit feedbacks described in the present invention. DETAILED DESCRIPTION

[0062] Now, in conjunction with the accompanying drawings of the specification, a specific implementation method of the multiple-pair ranking recommendation algorithm based on multiple implicit feedbacks described in the present invention is described in detail.

[0063] Example

[0064] The multiple-pair ranking recommendation algorithm based on multiple implicit feedbacks, such as Figure 1 As shown, the following steps are included:

[0065] Step 1. Data collection and processing stage: obtaining user behavior data from business websites and processing them;

[0066] Step 2. Recommendation algorithm processing stage: use the processed data to train the recommendation algorithm model and generate a recommendation model;

[0067] Step 3. In the recommendation list stage, the recommendation model is used to predict the user's ratings of items, and the item ratings are sorted, and finally a recommendation list is generated for each user.

[0068] As a specific embodiment, step 1, data collection preprocessing stage, obtains user behavior data from the business website and processes it, specifically including:

[0069] Step 1.1 Obtain the behavior data of cooperative users from the e-commerce website, process the data, filter and clean it to obtain implicit feedback data, which includes purchase data, click data and add to shopping cart data;

[0070] Step 1.2 Generate user-item matrix data based on purchase data, click data, and add-to-cart data, improve the item popularity algorithm based on the algorithm, and use it as a training set; and mine items that users do not like as auxiliary data;

[0071] Step 1.3 In order to make the user's preference assumption more reasonable, the popularity algorithm proposed by Steck H. is introduced to mine items that users do not like. The popularity of an item is defined as the number of users who interact with it. The more users interact with an item, the more popular the item is, and the higher the popularity of the item is. Considering multiple implicit feedbacks, that is, not only considering the impact of target behavior data on item popularity, but also considering the impact of auxiliary behavior data on item popularity, the popularity threshold ε is set p , we get a low-popularity item set I lp , as shown in formula (1):

[0072] I lp ={i∈I - |N i <ε p}......(1),

[0073] Among them, N i represents the number of users who have behavior on item i, when ε p When set to 1, find items with less than 1 user interaction and assign them to I lp For each user, the set of disliked items is obtained by the following formula (2):

[0074]

[0075] Based on auxiliary data, the items are further classified into the target behavior item set Add to cart item set Click Behavior Item Set Disliked Items and the remaining unknown set of items

[0076] Step 1.4 uses implicit feedback data and auxiliary data as training sets:

[0077] Sort the user's preferences for different items, where the items come from purchase data, click data, add to cart data, and auxiliary data, respectively. In this case, indicates that the user prefers purchased items over unknown items, the user prefers clicked items over unknown items, the user prefers added to cart items over unknown items, the user prefers purchased items over auxiliary items, the user prefers clicked items over auxiliary items, and the user prefers added to cart items over auxiliary items;

[0078] As a specific embodiment, step 2, using the user behavior data processed in the data collection and processing stage to train the recommendation algorithm and generate a recommendation model, includes:

[0079] Step 2.1: Combine the preference ranking pairs with the matrix decomposition model and the BPR model respectively, and initialize the model parameters. The parameters of the initialized model include: model parameter θ, learning rate η, feature vector dimension d, and regularization coefficient α;

[0080] Step 2.2 uses the sigmoid function to approximate the probability function P(·), and introduces the logarithm to reduce the complexity of the calculation, and finally obtains the objective function of the algorithm;

[0081] Step 2.3 uses stochastic gradient descent to learn the optimal value of the model parameter θ, traverses each preference ranking pair in the data set, and iterates the model and its parameters using the stochastic gradient descent algorithm;

[0082] Through the cyclic gradient descent algorithm, convergence iteration is achieved and the recommendation model is generated, that is, the optimal parameter setting of the proposed model is obtained through training.

[0083] As a specific embodiment, step 3, the recommendation list stage, uses the recommendation model to predict and sort the items based on the user behavior data and auxiliary data, specifically including:

[0084] Substitute user behavior data and auxiliary data into the recommendation model to generate a personalized recommendation list for each user: The parameters of the recommendation model are u, i, v, c, j, k, which are expressed as:

[0085] u is the user;

[0086] i is the item purchased by user u;

[0087] v is the item clicked by user u;

[0088] c is the item added to the shopping cart by user u;

[0089] j is an unknown item to user u;

[0090] k is the item that user u dislikes;

[0091] The following 6-tuple formula (3) is used to express the user's behavior preference of liking or disliking to buy items, click item v, and add items to the shopping cart relative to unknown items and disliked items:

[0092]

[0093] In formula (3), the 6-tuple formula indicates that user u prefers to buy item i relative to unknown item j, user u prefers to click item v relative to unknown item j, user u prefers to add item c to the shopping cart relative to unknown item j, user u prefers to buy item i relative to disliked item k, user u prefers to click item v relative to unknown item k, and user u prefers to add item c to the shopping cart relative to unknown item k;

[0094] in: is the set of items for purchase behavior, is the set of items for user click behavior, The set of items that users add to their shopping carts. For the set of disliked items, is the unknown set of items;

[0095] The sigmoid function is used to approximate the probability function P(·), and the logarithm is introduced to reduce the complexity of the calculation. Finally, the objective function of the recommendation algorithm is obtained as follows (5):

[0096]

[0097] Where: R(θ) is a regularization term used to prevent the algorithm from overfitting, as shown in the following formula (6):

[0098] λ u and λ v are the regularization parameters for users and items respectively:

[0099]

[0100] Here, the matrix decomposition framework is adopted, with a total of 6 algorithm parameters, the user feature vector U uand item feature vector V i 、V v 、V c 、V j and V k , where i, v, c, j and k come from and Set, θ is used to represent the parameters used in MPIF, and the specific parameters are shown in formula (7):

[0101]

[0102] The algorithm parameters are learned by using the stochastic gradient descent method. The update rule for each variable is as follows:

[0103]

[0104] Where η is the learning rate, which is used to control the step size of each iteration and the gradient of each variable, as shown in the following formula (9):

[0105]

[0106] The present invention is not limited to the above-mentioned embodiments. Without departing from the essential content of the present invention, any deformation, improvement, and substitution that can be thought of by those skilled in the art shall fall within the protection scope of the claims of the present invention.

Claims

1. A multi-pair ranking recommendation algorithm based on multiple implicit feedbacks, including: Step 1. Data collection and processing stage: obtaining user behavior data from business websites and processing them; Step 2. Recommendation algorithm processing stage: use the processed data to train the recommendation algorithm model and generate a recommendation model; Step 3. In the recommendation list stage, the recommendation model is used to predict the user's ratings of items, and the item ratings are sorted, and finally a recommendation list is generated for each user.

2. According to the multiple-pair ranking recommendation algorithm based on multiple implicit feedbacks as claimed in claim 1, the data collection and processing stage in step 1, obtaining user behavior data from the business website and processing it, specifically includes: Step 1.1 Obtain the behavior data of cooperative users from the e-commerce website, process the data, filter and clean it to obtain implicit feedback data, which includes purchase data, click data and add to shopping cart data; Step 1.2 Generate user-item matrix data based on purchase data, click data, and add-to-cart data, improve the item popularity algorithm based on the algorithm, and use it as a training set; and mine items that users do not like as auxiliary data; Step 1.3 To make the user preference assumption more reasonable, the popularity algorithm proposed by Steck H. is introduced to mine items that users do not like. The popularity of an item is defined as the number of users who interact with it. The more users interact with an item, the more popular the item is, and the higher the popularity of the item is. Considering multiple implicit feedbacks, that is, not only considering the impact of target behavior data on item popularity, but also considering the impact of auxiliary behavior data on item popularity, the popularity threshold ε is set. p , we get a low-popularity item set I lp , as shown in formula (1): I lp ={i∈I - |N i <e p }......(1), Among them, N i represents the number of users who have behavior on item i, when ε p When set to 1, find items with less than 1 user interaction and assign them to I lp For each user, the set of disliked items is obtained by the following formula (2): Based on auxiliary data, the items are further classified into the target behavior item set Add to cart item set Click Behavior Item Set Disliked Items and the remaining unknown set of items Step 1.4 uses implicit feedback data and auxiliary data as training sets: Sort the user's preferences for different items, and the items come from purchase data, click data, add to cart data and auxiliary data respectively. Among them, indicates that the user prefers purchased items over unknown items, the user prefers clicked items over unknown items, the user prefers added to cart items over unknown items, the user prefers purchased items over auxiliary items, the user prefers clicked items over auxiliary items, and the user prefers added to cart items over auxiliary items.

3. According to the multiple-pair ranking recommendation algorithm based on multiple implicit feedbacks as claimed in claim 1, the recommendation algorithm processing stage in step 2 uses the processed data to train the recommendation algorithm model and generate the recommendation model, specifically including: Step 2.1: Combine the preference ranking pairs with the matrix decomposition model and the BPR model respectively, and initialize the model parameters. The parameters of the initialized model include: model parameter θ, learning rate η, feature vector dimension d, and regularization coefficient α; Step 2.2 uses the sigmoid function to approximate the probability function P(·), and introduces the logarithm to reduce the complexity of the calculation, and finally obtains the objective function of the algorithm; Step 2.3 uses stochastic gradient descent to learn the optimal value of the model parameter θ, traverses each preference ranking pair in the data set, and iterates the model and its parameters using the stochastic gradient descent algorithm; Through the cyclic gradient descent algorithm, convergence iteration is achieved and the recommendation model is generated, that is, the optimal parameter setting of the proposed model is obtained through training.

4. According to the multiple-pair ranking recommendation algorithm based on multiple implicit feedbacks as claimed in claim 1, in step 3, the user behavior data and auxiliary data are substituted into the recommendation model to generate a personalized recommendation list for each user: the parameters of the recommendation model are u, i, v, c, j, k, which are respectively expressed as: u is the user; i is the item purchased by user u; v is the item clicked by user u; c is the item added to the shopping cart by user u; j is an unknown item to user u; k is the item that user u dislikes; The following 6-tuple formula (3) is used to express the user's behavior preference of liking or disliking to buy items, click item v, and add items to the shopping cart relative to unknown items and disliked items: In formula (3), the 6-tuple formula indicates that user u prefers to buy item i relative to unknown item j, user u prefers to click item v relative to unknown item j, user u prefers to add item c to the shopping cart relative to unknown item j, user u prefers to buy item i relative to disliked item k, user u prefers to click item v relative to unknown item k, and user u prefers to add item c to the shopping cart relative to unknown item k; in: is the set of items for purchase behavior, is the set of items for user click behavior, The set of items that users add to their shopping carts. For the set of disliked items, is the unknown set of items; The sigmoid function is used to approximate the probability function P(·), and the logarithm is introduced to reduce the complexity of the calculation. Finally, the objective function of the recommendation algorithm is obtained as follows (5): Where: R(θ) is a regularization term used to prevent the algorithm from overfitting, as shown in the following formula (6): λ u and λ v are the regularization parameters for users and items respectively: Here, the matrix decomposition framework is adopted, with a total of 6 algorithm parameters, the user feature vector U u and item feature vector V i 、V v 、V c 、V j and V k , where i, v, c, j and k come from and Set, θ is used to represent the parameters used in MPIF, and the specific parameters are shown in formula (7): The algorithm parameters are learned by using the stochastic gradient descent method. The update rule for each variable is as follows: Where η is the learning rate, which is used to control the step size of each iteration and the gradient of each variable, as shown in the following formula (9):

Images