Recommendation method and device based on user group sampling, equipment and storage medium

Abstract

The application discloses a recommendation method and device based on user group sampling, equipment and storage medium. The recommendation method is based on the implicit feedback data of the user obtained from the internet platform, the item entropy is modeled on the dynamically updated model, the global rating information of the item and the common interaction behavior of the user are fully utilized in optimization, and a more reliable similar user group is sampled. The pair group preference paradigm is used to learn the user group preference, and the potential favorite items of the user are mined, so that the recommendation effect of the improved model is improved, and the high interpretability and algorithm universality are also achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning and recommendation system, in particular, to a recommendation method and device based on user group sampling, equipment and storage medium. BACKGROUND

[0002] Collaborative filtering has been widely used in recommendation systems, which uses historical behavior data to predict user preferences. In real-world scenarios, the collected data is often implicit feedback data, that is, only the observed interaction behavior between users and items can be accessed, such as "likes" on social media or "purchases" on shopping websites. Implicit feedback data does not involve numerical ratings and cannot directly reflect the preference degree of users. A large amount of feedback information that is not observed by commercial platforms cannot be simply understood as negative feedback of users on these items. Among them, the pair-wise method is one of the mainstream methods for processing implicit feedback, which focuses on the relative order between independent user-item pairs and tries to maximize the possibility of pair-wise preference. Bayesian personalized ranking is one of the representative methods, and many pair-wise ranking methods are based on Bayesian personalized ranking and have been improved and proven effective in practice.

[0003] The characteristics of implicit feedback data bring great difficulties to improve the recommended algorithm to accurately learn user preferences. Although the above pair-wise ranking method is effective in solving the problem of implicit feedback, most of these methods are affected by the assumption that users are independent of each other, which contradicts the fact that users are always related in reality, making it impossible for the recommendation system to fully exploit the rich interaction information between users. Bayesian personalized ranking based on group preference introduces similar user groups to relax the assumption of user independence, providing a paradigm for pair-wise ranking methods to learn group preferences. However, when sampling similar user groups, it only focuses on limited similar interaction behaviors. Considering that implicit feedback data is usually very sparse, the reliability of the sampled user groups is likely to be low. There are many attempts to model user similarity that provide a reference for sampling similar users, but most of them focus on explicit ratings of users on items. In the implicit feedback scenario, users have the same attitude towards all the items they have interacted with, making it difficult to directly model the differences in user preferences, making it more difficult to sample similar user groups, and further limiting the accurate learning of user preferences by the recommendation system. SUMMARY

[0004] The purpose of the present application is to provide a recommendation method based on user group sampling, which can construct more reliable similar user groups, introduce more rich and reliable user interaction information for personalized recommendation systems, improve the learning ability of user group preferences, and improve the service quality of personalized recommendation provided by the recommendation system for users.

[0005] In order to achieve the above object, the first aspect of the present application provides a recommendation method based on user group sampling, which comprises:

[0006] Step 1, data collection, obtaining implicit feedback data of users from online platforms on the Internet;

[0007] Step 2, data preprocessing, dividing training data and test data according to the implicit feedback data obtained in step 1, optimizing user and item features by using a stochastic gradient descent algorithm and combining training data, learning a pair-wise ranking target based on random initialization or introducing group preference, obtaining implicit vector representation of users and implicit vector representation of items under the current model;

[0008] Step 3, similarity matrix calculation, calculating score entropy for each item based on the predicted score of all users to the item, modeling the similarity between users based on the predicted score difference of the items interacted by the users, and calculating the similarity matrix between users by combining the score entropy of the items;

[0009] Step 4, user group sampling, in each update, a user is randomly selected, and the most similar specified number of users to the user are selected according to the pre-calculated similarity matrix to form a user group;

[0010] Step 5, combining different target sampling corresponding number of positive and negative item samples, forming a ranking pair based on group preference to continue optimization, according to the final learning result of the model, according to the predicted preference degree of users to items, ranking the top several items as the items that users may like and recommending them to users.

[0011] Preferably, step 1 comprises:

[0012] Implicit feedback data is obtained from online platforms on the Internet by using a crawler technology, wherein,

[0013] For different online platforms, different data is collected;

[0014] For online social platforms, the attention records of users to other users are crawled;

[0015] For online media platforms, the browsing records of users to the media are crawled;

[0016] For online e-commerce platforms, the purchase records of users to goods are obtained according to open source data;

[0017] According to the collected data, a user-item matrix is constructed, assuming that the user is u and the item is i, the element (u, i) in the user-item matrix records the operation record of the user u to the item i, if the corresponding element is 1, it means that the user has interacted with the item, and the user has positive feedback to the item.

[0018] Preferably, step 2 comprises:

[0019] Using denotes a set of users, using denotes a set of items; where n and m denote the number of users and items respectively; for each user u e U, a positive feedback has been expressed on the item set To model the user preference, a matrix factorization model is used, which predicts the user rating for an item by where f ui denotes the predicted rating of item i by user u, U u· denotes the latent vector describing the feature of user u, V i. denotes the latent vector describing the feature of item i; the goal of recommendation is to rank the items that the user has not interacted with in a personalized item list for each user in descending order of the predicted rating of the user for the item.

[0020] Preferably, in step 2, Bayesian personalized ranking BPR is introduced, BPR assumes that the user u prefers the item i that has been interacted with to the item j that has not been interacted with, that is, a relative order is optimized, so that the ranking of the item i by the user u is more in front of the ranking of the item j; the pair-wise preference paradigm of BPR is:

[0021] f BPR = f ui - f uj

[0022] where,

[0023] In view of the fact that in the real world users are always connected with each other, and the personalized preference is affected by the group effect, Bayesian personalized ranking based on group preference G-BPR is introduced to introduce group preference into the pair-wise ranking target; the group preference of the user group G for the item i is estimated by the individual preference of the users in G:

[0024] To better learn the individual preference and group preference, the two are linearly combined to obtain the fused preference f Gui = p f Gi + (1-p) f ui ; where, p is a trade-off parameter;

[0025] In view of the fact that the fused preference f Gui of the item i is more than the preference f uj of the user u on the item that has not been interacted with, the pair-wise preference order optimization target of the group preference is introduced:

[0026] f​G-BPR = f Gui -f uj ;

[0027] Based on the above optimization objective, maximizing a maximum likelihood function of the optimization pair-wise ranking objective as an objective function, the model based on matrix decomposition, by means of stochastic gradient descent, the user feature hidden vector and the item feature hidden vector are optimized.

[0028] Preferably, step 3 comprises:

[0029] For two specific users u and v, using I u,v representing the set of items they have interacted with, the entropy E(i) of the item rating based on the overall user's score on the item is introduced into the similarity measure, and a reliable user group with similar preferences is established based on the entropy;

[0030] In the collaborative filtering problem, for a specific item i, the entropy is used to describe the possibility of different users giving different ratings to it; if the attitude of the user to the item is expressed in different degrees from l min to l max , the rating label of item i is represented by l i , the rating entropy E(i) of item i is calculated as:

[0031]

[0032] Where, if l represents the rating label of item i from user u, the possibility of the item receiving a rating of l is:

[0033]

[0034] Based on this, the entropy is applied to the Pearson correlation COR, which quantifies the strength of the correlation between two variables; when the value is 1, it represents a completely positive linear relationship, when the value is -1, it represents a completely negative linear relationship, and when the value is 0, it represents no linear relationship; COR measures the linear relationship between variables using the deviation from the average rating of the user, let I u,v represent the set of items that users u and v represent positive feedback, represent the average predicted score of the items in I u,v by user u; then, the similarity measure to be used combines the COR between users u and v, and the COR ent is represented as:

[0035]

[0036] Preferably, step 4 comprises:

[0037] When using the stochastic gradient descent method for optimization, each round of update requires randomly selecting a user u from the user set as the main target user for optimization, building a user group for users with similar sampled preferences, first selecting an array containing the similarities between user u and other users from the similarity matrix, sorting them according to similarity, and selecting a specified number of the most similar users to form the user group G.

[0038] A second aspect of the present invention provides a personalized recommendation device based on user group sampling, the personalized recommendation device comprising:

[0039] A data collection module is used to obtain implicit feedback data of users from an Internet online platform;

[0040] The data preprocessing module is used to divide the implicit feedback data obtained by the data collection module into training data and test data. It uses the stochastic gradient descent algorithm combined with the training data to optimize user and item features. Based on the random initialization or the introduction of group preference, the level ranking objective learning is used to obtain the latent vector representation of users and items under the current model.

[0041] The similarity matrix calculation module is used to calculate the rating entropy for each item based on the predicted ratings of all users for the items. It then models the similarity between users based on the predicted rating differences of items that have been interacted with by two users. The similarity matrix between users is obtained by weighted calculation based on the rating entropy of the items.

[0042] The user group sampling module is used to randomly select a user in each update and select a specified number of users who are most similar to the user according to the pre-calculated similarity matrix to form a user group;

[0043] The recommendation module is used to combine the corresponding number of positive and negative item samples for different target sampling to form a ranking pair based on group preference for further optimization. Based on the final learning results of the model and the predicted user preference for items, several top-ranked items are recommended to the user as items that the user may like.

[0044] The third aspect of the present invention provides a personalized recommendation device based on user group sampling, which includes: a memory and at least one processor, wherein instructions are stored in the memory, and the memory and the at least one processor are interconnected through a line; the at least one processor calls the instructions in the memory so that the personalized recommendation device performs the steps of the above-mentioned recommendation method based on user group sampling.

[0045] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the above-mentioned recommendation method based on user group sampling are implemented.

[0046] According to the above technical solution, the present invention uses an entropy-based user group sampling method to sample user groups with similar preferences from an implicit feedback peer ranking model. Compared to other methods, this method incorporates global rating information, enabling the construction of more reliable similar user groups. Furthermore, this method introduces richer and more reliable user interaction information into the personalized recommendation system, enhancing the learning of user group preferences and improving the quality of personalized recommendation services provided by the recommendation system to users.

[0047] Other features and advantages of the present invention will be described in detail in the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the following detailed description, they are used to explain the present invention but do not constitute a limitation of the present invention. In the accompanying drawings:

[0049] Figure 1 It is a schematic diagram of user group sampling in the recommendation method based on user group sampling provided by the present invention. DETAILED DESCRIPTION

[0050] The following describes the specific embodiments of the present invention in detail with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only used to illustrate and explain the present invention and are not intended to limit the present invention.

[0051] A first aspect of the present invention provides a recommendation method based on user group sampling, the method mainly comprising:

[0052] Step 1: Obtain users' implicit feedback data from the Internet online platform.

[0053] Specifically, we first use crawler technology to obtain implicit feedback data from online internet platforms. Different data is collected for different online platforms. For online social platforms, we crawl users' attention records for other users; for online media platforms, we crawl users' browsing records of current media; and for online e-commerce platforms, we obtain users' purchase records of products based on open source data. Based on the collected data, we construct a user-item matrix. Let's assume user u and item i. The element (,i) in the user-item matrix records the actions taken by user u on item i. If the corresponding element is 1, it means that the user has interacted with the item and has provided positive feedback on the item.

[0054] Step 2: Divide the implicit feedback data obtained in step 1 into training data and test data, use the stochastic gradient descent algorithm in combination with the training data to optimize user and item features, and learn the level ranking objective based on random initialization or the introduction of group preference to obtain the latent vector representation of users and items under the current model.

[0055] Specifically, we first give a formal expression of the problem, using Represents a user collection, using Represents the item set; where n and m represent the number of users and the number of items respectively. For each user u∈U, In order to model user preferences, a matrix decomposition model is used, which can predict the user's rating of the item by Calculated. Among them, f ui represents the predicted rating of user u for item i, U u. represents the latent vector describing the characteristics of user u, V i. Represents the latent vector describing the features of item i. The goal of recommendation is to create a set of items that have never been interacted with. In [1], a personalized list of items is generated for each user, sorted in descending order by their predicted item scores. The following example uses Bayesian personalized ranking with group preferences to illustrate how the model optimizes and learns based on data:

[0056] BPR is a paradigm for solving the implicit feedback problem. It assumes that user u prefers items i that have been interacted with to items j that have not been interacted with. In other words, it optimizes a relative order so that user u ranks item i higher than item j. The pairwise preference paradigm of BPR can be written as follows:

[0057] f BPR =f ui -f uj

[0058] in,

[0059] However, in the real world, users are always connected to each other, and their individual preferences are often affected by group effects. Group-Preference Bayesian Personalized Ranking (G-BPR) extends BPR by introducing group preferences into the pairwise ranking objective. The group preference of a user group G for item i can be estimated from the individual preferences of users in G: In order to better learn individual preferences and group preferences, it is proposed to linearly combine the two to obtain the fusion preference f Gui =ρf Gi +(1-\rho)f ui; where p is a trade-off parameter. And here we propose to incorporate the fusion preference f Gui of item i into the similarity measure between users u and v, which is likely to be greater than the preference f uj of user u on the items that they have not interacted with.

[0060] f G-BPR = f Gui - f uj .

[0061] The above is an example of incorporating group preference into many pair-wise ranking methods based on the assumption of user independence. Such methods are equally applicable to other pair-wise ranking methods. Based on the above optimization objective, the maximum likelihood function of maximizing an optimization pair-wise ranking objective is taken as the objective function, and the matrix decomposition-based model is optimized by the way of stochastic gradient descent to optimize the user feature hidden vector and the item feature hidden vector.

[0062] Step 3, calculate the score entropy of each item based on the predicted scores of all users on the item, and model the similarity between users based on the predicted score difference of the items that are interacted by both users, and combine the score entropy of the item to calculate the similarity matrix between users.

[0063] For two specific users u and v, most of the previous indicators for measuring their similarity are focused on the items that they have interacted with. Use I u,v to represent the set of items that they have interacted with, which is likely to be a very small set, because in real-world commercial recommendation systems there are usually a huge number of unobserved items. Therefore, based on the previous indicators, the calculated similarity between two users is usually not reliable enough. In view of this problem, the entropy E(i) of the item score can be calculated based on the scores of all users on the item, and it is introduced into the similarity measure to more accurately measure the similarity and establish a reliable group of users with similar preferences.

[0064] Entropy is usually defined in information theory as the probability of occurrence of a discrete random event. In the problem of collaborative filtering, for a specific item i, entropy can be used to describe the possibility of different users producing different evaluations of it. If the attitude of users towards the item is expressed in different degrees from l min to l max , and l i represents the rating label of item i, then the rating entropy E(i) of item i can be calculated as:

[0065]

[0066] where if we denote the rating label of item i from user u as l, then the likelihood of item i receiving a rating of l is:

[0067]

[0068] Based on the above, we apply entropy to one of the most representative similarity measures, Pearson Correlation (COR). COR can quantify the strength of the association between two variables, with a value of 1 indicating a perfect positive linear relationship, a value of -1 indicating a perfect negative linear relationship, and a value of 0 indicating no linear relationship. It should be noted that other traditional similarity measures, such as COS, MSD, etc., can also use the same principle. COR measures the linear relationship between variables using the deviation from the average rating of the user. Let I u,v denote the set of items that users u and v indicate positive feedback on, denote the average predicted rating of item i by user u. Then, the similarity measure to be adopted combines the COR between users u and v, denoted as COR u,v , which is formally defined as: ent

[0069]

[0070] As can be seen, the entropy of the overall user rating of an item reflects a global information, and thus can help us better understand the behavior of users. The entropy-weighted similarity can be used to find similar users to form user groups. The main idea of using entropy to measure user similarity in the user group sampling strategy provided in the present application can be described as follows: the behavior similarity of two users on high-entropy items should be higher than that on low-entropy items. Similarly, when the entropy of an item is low, the rating difference between two users is large, which should still result in a lower preference difference. This can be interpreted in two ways: first, for those items with high entropy, the user's attitude towards the item is mixed, so the item is likely to be a personalized item. The consistency of the behavior of two users on these items indicates a higher likelihood of their similarity; second, items with low entropy mean that users have similar attitudes towards the item. Therefore, the probability of users showing similarity on low-entropy items should be lower than that on high-entropy items.

[0071] In the present application, in order to calculate the user similarity matrix, first, the rating entropy is calculated for each item in the item set. For each item i, the current predicted rating f u,i ​and convert it into the nearest integer score label, and calculate the entropy E(i) of the item according to the probability of the predicted score falling on different score labels; then, cycle the user set, and calculate the similarity between each pair of users according to the entropy-weighted user similarity measure, and store it in the user similarity matrix.

[0072] Based on the user similarity matrix calculated above, similar users can be easily selected to form a user group. Moreover, the sampling strategy of this method not only relies on the common interaction of users on items, but also refers to the overall behavior of users for sampling, which makes the user group introduce more rich interactions between users for training.

[0073] Step 4, user group sampling, in each update, a user is randomly selected, and the most similar specified number of users to the user are selected according to the pre-calculated similarity matrix to form a user group.

[0074] Specifically, when using the method of stochastic gradient descent for optimization, each round of update first needs to randomly select a user from the user set as the main target user u for optimization. In order to sample similar preference users to build a user group, the array containing the similarity of user u and other users is first selected from the similarity matrix, and the specified number of most similar users is selected to form a user group G according to the similarity.

[0075] Step 5, combine the corresponding number of positive and negative item samples of different target sampling to form a ranking based on group preference for further optimization, and according to the final learning result of the model, the top ranked items are recommended to the user as the items that the user may like according to the predicted preference of the user for the items.

[0076] In summary, referring to Figure 1 which shows the sampling process of the entropy-based user group sampling algorithm in the implicit feedback scenario. In order to find similar users and form a user group, the entropy of each item is first calculated, which is different from the explicit feedback problem of directly calculating the entropy using numerical scores. People have fixed several levels for the score of an item. In order to better utilize the preference information of users for different items, the method selects the model to dynamically update the predicted score as a soft label, and further converts the predicted score into discrete multi-labels. Here, the continuous score f uiThe conversion to the nearest integer is taken as the label l. Based on the probability under discrete multi-label, the rating entropy of an item can be derived from the predicted scores of all users on the item according to the formula of entropy. When the entropy is ready, it can be used as a weight reference to calculate the similarity between users, such as COR, etc. As can be seen from the figure, the similarity calculation is based on the set of items that the two users have interacted with and the global evaluation entropy of each item. Then, for each update, first randomly sample a user u (step 3), an interacted item i and a non-interacted item j, then sort the user list according to the relative similarity of the user to u, and pick the most similar user (step 4) to form a user group G, where u∈G. Based on this, a new pairwise group preference optimization objective can be successfully constructed. Further, the user's preference is accurately learned in the recommendation algorithm, and the goal of providing personalized recommendation services for the user is achieved.

[0077] In addition, the second aspect of the present application provides a personalized recommendation device based on user group sampling, which comprises:

[0078] A data collection module for obtaining implicit feedback data of a user from an Internet online platform;

[0079] A data preprocessing module for dividing the implicit feedback data obtained by the data collection module into training data and test data, optimizing the user and item features by combining the training data with the random gradient descent algorithm, learning the implicit vector representation of the user and the implicit vector representation of the item based on the pair-wise ranking objective learning with random initialization or introducing group preference under the current model;

[0080] A similarity matrix calculation module for calculating the rating entropy of each item based on the predicted scores of all users on the item, and modeling the similarity between users based on the predicted score difference of the items that the two users have interacted with, and calculating the similarity matrix between users by combining the rating entropy of the items;

[0081] A user group sampling module for randomly selecting a user in each update, and selecting the most similar specified number of users to the user according to the pre-calculated similarity matrix to form a user group;

[0082] A recommendation module for combining different target sampling corresponding number of positive and negative item samples to form a ranking pair based on group preference for further optimization, and according to the final learning result of the model, recommending the top ranked items to the user as the items that the user may like according to the predicted preference degree of the user to the items.

[0083] Using the device, a more reliable similar user group can be constructed, richer and reliable user interaction information can be introduced into the personalized recommendation system, the learning ability for user group preferences can be improved, and the service quality of the personalized recommendation provided by the recommendation system for users can be improved.

[0084] Meanwhile, the third aspect of the present application further provides a personalized recommendation device based on user group sampling, which comprises a memory and at least one processor, the memory has instructions stored therein, and the memory and the at least one processor are interconnected through a circuit; the at least one processor invokes the instructions in the memory, so that the personalized recommendation device performs the steps of the above-mentioned recommendation method based on user group sampling.

[0085] Through the device, a more reliable similar user group can be constructed, richer and reliable user interaction information can be introduced into the personalized recommendation system, the learning ability for user group preferences can be improved, and the service quality of the personalized recommendation provided by the recommendation system for users can be improved.

[0086] Further, the fourth aspect of the present application further provides a computer readable storage medium, which has a computer program stored thereon, and the computer program is executed by a processor to realize the steps of the above-mentioned recommendation method based on user group sampling.

[0087] Through the computer readable storage medium, a more reliable similar user group can be constructed, richer and reliable user interaction information can be introduced into the personalized recommendation system, the learning ability for user group preferences can be improved, and the service quality of the personalized recommendation provided by the recommendation system for users can be improved.

[0088] The preferred embodiments of the present application are described in detail above with reference to the accompanying drawings, but the present application is not limited to the specific details in the above-described embodiments, and various simple modifications can be made to the technical solutions of the present application within the technical concept of the present application, and these simple modifications all belong to the protection scope of the present application.

[0089] In addition, it should be noted that each specific technical feature described in the above-described specific embodiments can be combined in any appropriate manner without contradiction, and in order to avoid unnecessary repetition, the present application will not further describe various possible combinations.

[0090] In addition, various different embodiments of the present application can also be combined in any manner, as long as they do not deviate from the technical concept of the present application, and they should also be considered as disclosed by the present application.

Claims

1. A recommendation method based on user group sampling, characterized in that: The method comprises: Step 1: Data collection: Obtain users’ implicit feedback data from online platforms. Step 2: Data preprocessing: Divide the implicit feedback data obtained in step 1 into training data and test data. Use the stochastic gradient descent algorithm combined with the training data to optimize user and item features. Learn the ranking objective based on random initialization or the introduction of group preferences to obtain the latent vector representation of users and items under the current model. Step 3: Calculate the similarity matrix. Calculate the rating entropy for each item based on the predicted ratings of all users. Then, model the similarity between users based on the predicted rating difference of items that have been interacted with by two users. Combined with the item rating entropy, a weighted calculation is performed to obtain the similarity matrix between users. Step 4: User group sampling: In each update, a user is randomly selected, and a specified number of users who are most similar to the user are selected according to the pre-calculated similarity matrix to form a user group; Step 5: Combine the corresponding number of positive and negative item samples for different target sampling to form a group preference-based ranking pair for further optimization. Based on the final learning results of the model and the predicted user's preference for items, the top-ranked items are recommended to the user as items that the user may like; Step 2 includes: use Represents a user collection, using Represents a set of items; where n and m represent the number of users and the number of items respectively; for each user u∈U, The positive feedback is expressed above; to model user preferences, a matrix decomposition model is used, which predicts the user's rating of the item by Calculated; where f ui represents the predicted rating of user u for item i, U u· represents the latent vector describing the characteristics of user u, V i· Represents the latent vector describing the characteristics of item i; the target of recommendation is the set of items that have never been interacted with In the above example, a personalized item list is obtained for each user by sorting the items in descending order according to the user's predicted scores; In step 2, Bayesian personalized ranking (BPR) is introduced. BPR assumes that user u prefers interacted item i to uninteracted item j. In other words, it optimizes a relative order so that user u ranks item i higher than item j. The pairwise preference paradigm of BPR is: f BPR =f ui -f uj in, In view of the fact that users in the real world are always connected with each other and their personalized preferences are affected by the group effect, we introduce the Bayesian personalized ranking (G-BPR) based on group preference to introduce group preference into the pairwise ranking objective. The group preference of user group G for item i is estimated by the individual preferences of users in G: In order to better learn individual preferences and group preferences, the two are linearly combined to obtain the fusion preference f Gui =ρf Gi +(1-\rho)f ui ;in, ρ is a trade-off parameter; Fusion preference f for item i Gui It is likely to be greater than user u's preference f for items that have not been interacted with. uj , introducing the pairwise preference order optimization objective of group preference: f G-BPR =f Gui -f uj ; Based on the above optimization objectives, a maximum likelihood function that optimizes the pairwise sorting objective is maximized as the objective function. Based on the matrix decomposition model, the user feature latent vector and the item feature latent vector are optimized by stochastic gradient descent.

2. The recommendation method based on user group sampling according to claim 1, characterized in that: Step 1 includes: The crawler technology is used to obtain implicit feedback data from the Internet online platform, among which, For different online platforms, collect corresponding different data; Crawling the records of users following other users on online social platforms; For online media platforms, crawl users’ browsing history of current media; Online e-commerce platforms obtain users' purchase records of products based on open source data; Based on the collected data, a user-item matrix is ​​constructed. Assume that the user is u and the item is i. The element (u, i) in the user-item matrix records the operation record of user u on item i. If the corresponding element is 1, it means that the user has interacted with this item and the user has generated positive feedback for this item.

3. The recommendation method based on user group sampling according to claim 1, characterized in that: Step 3 includes: For two specific users u and v, use I u,v Represents the set of items they have interacted with together, introduces the entropy E(i) of the item rating calculated based on the scores of all users on the items into the similarity measurement, and establishes reliable user groups with similar preferences based on entropy; In collaborative filtering problems, for a specific item i, entropy is used to describe the possibility that different users will have different evaluations of it; if the user's attitude towards the item is expressed in different degrees from l min to l max , use l i Represents the rating label of item i, and the rating entropy E(i) of item i is calculated as: Among them, if l is used to represent the rating label of item i from user u, then the probability of the item receiving a rating of l is: Based on this, entropy is applied to the Pearson correlation COR, which quantifies the strength of the association between two variables. When the value is 1, it indicates a completely positive linear relationship, when the value is -1, it indicates a completely negative linear relationship, and when the value is 0, it indicates no linear relationship. COR uses the deviation from the average user rating to measure the linear relationship between variables. Let I u,v represents the set of items for which users u and v have positive feedback, Indicates user u's relationship with I u,v The average predicted score of the items in the statistic; Then, the similarity measure to be adopted combines the COR between users u and v, using COR ent Expressed as:

4. The recommendation method based on user group sampling according to claim 1, characterized in that: Step 4 includes: When using the stochastic gradient descent method for optimization, each round of update requires randomly selecting a user u from the user set as the main target user for optimization, building a user group for users with similar sampled preferences, first selecting an array containing the similarities between user u and other users from the similarity matrix, sorting them according to similarity, and selecting a specified number of the most similar users to form the user group G.

5. A personalized recommendation device based on user group sampling, characterized in that: The personalized recommendation device performs the steps of the recommendation method based on user group sampling according to any one of claims 1 to 4, including: A data collection module is used to obtain implicit feedback data of users from an Internet online platform; The data preprocessing module is used to divide the implicit feedback data obtained by the data collection module into training data and test data. It uses the stochastic gradient descent algorithm combined with the training data to optimize user and item features. Based on the random initialization or the introduction of group preference, the level ranking objective learning is used to obtain the latent vector representation of users and items under the current model. The similarity matrix calculation module is used to calculate the rating entropy for each item based on the predicted ratings of all users for the items. It then models the similarity between users based on the predicted rating differences of items that have been interacted with by two users. The similarity matrix between users is obtained by weighted calculation based on the rating entropy of the items. The user group sampling module is used to randomly select a user in each update and select a specified number of users who are most similar to the user according to the pre-calculated similarity matrix to form a user group; The recommendation module is used to combine the corresponding number of positive and negative item samples for different target sampling to form a ranking pair based on group preference for further optimization. Based on the final learning results of the model and the predicted user preference for items, several top-ranked items are recommended to the user as items that the user may like.

6. A personalized recommendation device based on user group sampling, characterized in that: The personalized recommendation device includes: a memory and at least one processor, the memory storing instructions, and the memory and the at least one processor being interconnected via a line; the at least one processor calls the instructions in the memory so that the personalized recommendation device performs the steps of the recommendation method based on user group sampling as described in any one of claims 1-4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the recommendation method based on user group sampling according to any one of claims 1 to 4 are implemented.

Images