Fair re-ranking recommendation method and system for candidate project pool based on project perspective

Abstract

The present invention relates to the technical field of recommendation system re-ranking, and provides a method and system for fairness re-ranking recommendation of a candidate project pool based on a project perspective. First, a data set for training is obtained; the data set for training is processed to obtain a bipartite graph of user-project interaction, a user initial embedding vector, and a project initial embedding vector, and optimized to obtain an optimized embedding vector representation; the optimized embedding vector representation is calculated to generate the user's initial preference score for the project and the initial user's preference ranking for the project, and multiple re-rankings are performed based on the initial preference score and the initial user's preference ranking for the project to obtain the user's preference ranking for the project, and a recommendation list is obtained. The method of the present invention provides a new sorting perspective, which has better performance than similar technologies in terms of sorting accuracy and fairness results, and has great advantages in operating efficiency. This method is superior to other existing similar methods in terms of recommendation performance.

Description

Technical Field

[0001] The present invention relates to the technical field of recommendation system rearrangement, and provides a method and system for fair rearrangement recommendation of a candidate project pool based on a project perspective. Background Art

[0002] With the development of recommendation systems, many methods have been proposed to improve the quality of recommendations. Generally speaking, many recommendation systems have implemented top-k recommendations, which include selecting the k items with the highest scores generated by a ranking function or algorithm for each user.

[0003] Mainstream recommendation systems have achieved great success by focusing on how to recommend items that users are interested in. However, in most systems, there is a long tail effect in the number of recommended items, which may lead to unfairness to certain items. The vast majority of promoted items are very popular, which may result in unpopular items having no chance to be recommended or effectively recommended. From the perspective of content providers, people also hope that all content provided by the platform can be displayed and appreciated by users, which is also one of the fairness issues in the recommendation system. Summary of the invention

[0004] Specifically, existing evaluation indicators usually evaluate the recommendation forms such as the exposure opportunities and position deviation of the project, but are separated from the measurement of accuracy and cannot measure the actual benefits brought by fairness. Among the fairness recommendation methods, there are mainly two types of methods: model-based and rearrangement-based. Among them, the rearrangement-based method only needs to be performed on the output recommendation results of the existing model, and has better transferability and flexibility than the model-based method. However, both the model-based and rearrangement-based fairness recommendation methods ignore the fairness benefits brought by the project perspective, resulting in the system's insufficient performance in terms of project fairness while maintaining the accuracy of the recommendation, and there is still a serious long-tail effect. Although some methods have achieved good results in the bilateral optimization of accuracy and exposure opportunities, they still do not consider the specific fairness benefits of a single project, resulting in some projects being treated fairly in terms of exposure, but not being improved in terms of fairness in terms of actual benefits. In addition, the time spent on the execution of the existing fairness methods in the recommendation system accounts for a large proportion, and fairness, as a long-term optimization indicator of the recommendation system, has a large room for optimization in terms of execution efficiency.

[0005] The present invention aims to solve at least one of the technical problems existing in the related art. To this end, the present invention provides a fair re-ranking recommendation method and system for a candidate project pool based on a project perspective, which can effectively mine long-tail projects in the recommendation system and improve the fairness benefits of the projects.

[0006] The present invention provides a method for fair rearrangement recommendation of a candidate project pool based on a project perspective, the steps comprising:

[0007] S1: Process the interaction data between users and projects to obtain a data set for training;

[0008] Processing the training data set to obtain a user-item interaction bipartite graph;

[0009] Construct the number of users and items to obtain the user initial embedding vector and the item initial embedding vector;

[0010] S2: Obtaining an optimized embedding vector representation through the user-item interaction bipartite graph, the user's initial embedding vector and the item's initial embedding vector;

[0011] S3: Calculate the optimized embedding vector representation to generate the user's initial preference score for the project, rank the initial preference scores, and obtain the initial user's preference ranking for the project;

[0012] S4: re-ranking is performed based on the initial preference scores and the initial user's preference rankings for the items to obtain a recommendation tendency ranking of the items for the users;

[0013] S5: re-ranking the user's recommendation tendency ranking of the items to obtain the user's preference ranking for the items and obtain a recommendation list.

[0014] According to a method for fair re-ranking and recommending a candidate project pool based on a project perspective provided by the present invention, step S1 comprises:

[0015] S11: acquiring interaction data between users and items from a public data set, taking items in which users and items interact in the interaction data between users and items as positive examples, and randomly sampling items in which users and items do not interact in the interaction data between users and items to obtain negative examples, so that the number of negative examples is the same as the number of positive examples, and the positive examples and negative examples constitute a data set for training;

[0016] S12: Convert the training data set into a matrix to obtain a user-item interaction bipartite graph, which is represented by a matrix A, ,in, is the set of all users to be recommended. is the candidate pool project set, is the set of real numbers;

[0017] S13: Vectorize the number of users and items in the interaction data of users and items to obtain the user's initial embedding vector and the item initial embedding vector ,in, , , d Indicates the dimension size of each embedding vector.

[0018] According to a method for fair re-ranking and recommending a candidate project pool based on a project perspective provided by the present invention, step S2 comprises:

[0019] S21: Through the user-item interaction bipartite graph, the user initial embedding vector and the item initial embedding vector, the graph convolution technology is used to learn the user initial embedding vector and the item initial embedding vector. The calculation formula is:

[0020]

[0021]

[0022]

[0023]

[0024]

[0025] in, For users to be recommended, For candidate projects, is the ordinal number of graph convolutions, represents the matrix transpose, is the user embedding vector representation without graph convolution, Embedding vector representation for items without graph convolution, Indicates The user embedding vector representation obtained by the graph convolution is Indicates The item embedding vector representation obtained by the graph convolution is For the matrix A The Laplace normalized matrix of is the diagonal matrix of the node degrees of the user graph, , is the diagonal matrix of node degrees of user and project graphs, ;

[0026] S22: Calculate and optimize the interaction loss by using the user embedding vector representation, the item embedding vector representation and the interaction information of the training dataset after graph convolution learning ,

[0027]

[0028]

[0029]

[0030] in, is the total number of graph convolution operations, is the average user embedding vector representation, is the average item embedding vector representation, represents the dataset used for training, Is a collection The variance of is the negative example subscript, is the sample represented by the average user embedding vector, is the sample of interactive positive examples in the average vector embedding representation, is a sample of uninteracted negative examples in the average item embedding vector representation;

[0031] S23: Perform multiple graph convolution learning on the user embedding vector representation and the item embedding vector representation, use the adaptive moment estimation optimizer to update the parameters of the user embedding vector representation and the item embedding vector representation, and go through multiple rounds of optimization of the interaction loss until the user embedding vector representation and the item embedding vector representation gradually converge to obtain the optimized embedding vector representation.

[0032] According to a method for fair re-ranking and recommending a candidate project pool based on a project perspective provided by the present invention, step S3 comprises:

[0033] S31: Calculate the user's initial preference score for the item through the optimized embedding vector representation ,in, ;

[0034] S32: Reduce the initial preference scores to obtain a preference score matrix and the retained items matrix ,in, , ,in, To retain the number, the number of retained numbers must be greater than or equal to the number of sorted numbers;

[0035] S33: Sorting the preference score matrix to obtain the initial user's preference ranking for the project.

[0036]

[0037] in, Indicates the initial user About Project The preference ranking of Indicated in Project The reverse ranking of Indicates user Reserved A collection of items.

[0038] According to a method for fair re-ranking and recommending a candidate project pool based on a project perspective provided by the present invention, step S4 comprises:

[0039] The recommendation tendency ranking of the project to the user is obtained by re-ranking the initial preference score and the initial user's preference ranking of the project.

[0040]

[0041] in, Display items i For users u The recommended tendency ranking, Indicated in In, users In ascending order, express Project i The set of users with preference scores, Represents the norm of the matrix.

[0042] According to a method for fair re-ranking and recommending a candidate project pool based on a project perspective provided by the present invention, step S5 comprises:

[0043] By adding the number of users who interacted with the project in the training dataset to the recommendation tendency ranking of the project to the user, the user's preference ranking for the project is obtained.

[0044]

[0045] in, Indicates user About Project The preference ranking of , indicating that Project 's ascending ranking.

[0046] According to a fair re-ranking recommendation method for a candidate project pool based on a project perspective provided by the present invention, when the number of projects in the recommendation list is less than the number required for sorting, the initial preference ranking of the user for the project is Select the most preferred item that is not recommended to fill in the gap.

[0047] The present invention also provides a candidate project pool fairness re-ranking recommendation system based on project perspective, comprising:

[0048] Acquisition module: Process the interaction data of users and projects to obtain a data set for training; process the data set to obtain a user-project interaction bipartite graph; construct the number of users and projects to obtain the user initial embedding vector and the project initial embedding vector;

[0049] Calculation module: obtaining an optimized embedding vector representation through the user-item interaction bipartite graph, the user's initial embedding vector and the item's initial embedding vector; calculating the optimized embedding vector representation to generate the user's initial preference score for the item;

[0050] Sorting module: rank the initial preference scores to obtain the initial user's preference ranking for the project;

[0051] Re-ranking is performed based on the initial preference scores and the initial user's preference rankings for items to obtain a recommendation tendency ranking of items for users;

[0052] The user's recommendation tendency ranking of the items is re-ranked to obtain the user's preference ranking for the items and obtain a recommendation list.

[0053] The above one or more technical solutions in the embodiments of the present invention have at least one of the following technical effects:

[0054] The present invention provides a candidate project pool fairness rearrangement recommendation method and system based on project perspective, which obtains the user's preference ranking for the project by calculating the user's preference score for the project, and obtains the recommendation tendency of the project to the user from the project perspective.

[0055] Compared with the existing rearrangement technology, the method of the present invention provides a new sorting perspective, has better performance in sorting accuracy and fairness results than similar technologies, and has great advantages in operating efficiency. This method is superior to other existing similar methods in recommendation performance.

[0056] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0058] Figure 1 The present invention is a flowchart of a method for fair re-ranking and recommending a candidate item pool based on an item perspective according to an embodiment of the present invention.

[0059] Figure 2 Schematic diagram of a fair re-ranking recommendation method for a candidate item pool based on an item perspective according to an embodiment of the present invention.

[0060] Figure 3 The fairness and accuracy results of the embodiment of the present invention and other methods on the fairness re-ranking task of dataset 1 are compared.

[0061] Figure 4 The fairness and accuracy results of the embodiment of the present invention and other methods on the fairness re-ranking task of dataset 3 are compared.

[0062] Figure 5 The structure block diagram of the fairness rearrangement recommendation system of the candidate item pool based on the item perspective according to an embodiment of the present invention.

[0063] Reference numerals:

[0064] 101. Collection module; 102. Calculation module; 103. Sorting module. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme in the present invention will be clearly and completely described below. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present invention. The following embodiments are used to illustrate the present invention, but cannot be used to limit the scope of the present invention.

[0066] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the embodiment of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0067] Example

[0068] Combination Figures 1 to 5 Describing the Invention:

[0069] See also Figure 1 , Figure 1This is a flow chart of a method for fair re-ranking and recommending a candidate item pool based on an item perspective according to an embodiment of the present invention:

[0070] S1: Process the interaction data between users and projects to obtain a data set for training;

[0071] Processing the training data set to obtain a user-item interaction bipartite graph;

[0072] Construct the number of users and items to obtain the user initial embedding vector and the item initial embedding vector;

[0073] S2: Obtaining an optimized embedding vector representation through the user-item interaction bipartite graph, the user's initial embedding vector and the item's initial embedding vector;

[0074] S3: Calculate the optimized embedding vector representation to generate the user's initial preference score for the project, rank the initial preference scores, and obtain the initial user's preference ranking for the project;

[0075] S4: re-ranking is performed based on the initial preference scores and the initial user's preference rankings for the items to obtain a recommendation tendency ranking of the items for the users;

[0076] S5: re-ranking the user's recommendation tendency ranking of the items to obtain the user's preference ranking for the items and obtain a recommendation list.

[0077] Specifically, in step S1:

[0078] S11: acquiring interaction data between users and items from a public data set, taking items in which users and items interact in the interaction data between users and items as positive examples, and randomly sampling items in which users and items do not interact in the interaction data between users and items to obtain negative examples, so that the number of negative examples is the same as the number of positive examples, and the positive examples and negative examples constitute a data set for training;

[0079] Step S12 constructs a user-item interaction bipartite graph based on the interaction information in the training dataset, which is represented by the matrix .in, is the set of all users to be recommended. is the candidate pool project set, is the set of real numbers.

[0080] Step S13 constructs an initial embedding vector lookup table based on the number of users and items to represent the user's initial embedding vector and the item initial embedding vector .in, , , dIndicates the dimension size of each embedding vector.

[0081] In step S2, S21: Use graph convolution technology to learn the embedding vector representation of users and projects through the user-project interaction bipartite graph, the user's initial embedding vector and the project's initial embedding vector. Based on the user-project interaction bipartite graph constructed in step S1 and the user-project's initial embedding vector, use graph convolution technology to learn the embedding vector representation of users and projects. The calculation method is:

[0082]

[0083]

[0084] in, For users to be recommended, For candidate projects, is the ordinal number of graph convolutions, represents the matrix transpose, Indicates l The user vector representation obtained by the graph convolution is Indicates l The item vector representation obtained by the graph convolution is, in particular: , The initial vector is used to eliminate the matrix A The influence of the degree of the middle graph nodes, For the matrix A The Laplace normalized matrix of is calculated as:

[0085]

[0086] in, is the diagonal matrix of the node degrees of the user graph, , is the diagonal matrix of node degrees of user and project graphs, .

[0087] For each user or project, the number of projects the user interacts with is the degree of the user node, and the number of users the project interacts with is the degree of the project node. In order to enhance the vector representation performance of users and projects, the vector representation obtained by each graph convolution is averaged to obtain the average user and project embedding vector representation:

[0088]

[0089]

[0090] in, L is the total number of graph convolution operations, is the average user embedding vector representation, is the average project embedding vector representation. At this point, we can get the vector representation of users and projects after graph convolution learning and updating.

[0091] Step S22 calculates the interaction loss based on the average user vector representation and item vector representation after graph convolution learning and the interaction information in the training data , the specific method is as follows:

[0092]

[0093] in, represents the dataset used for training, Is a collection The variance of is the sample represented by the average user embedding vector, is the sample of interactive positive examples in the average vector embedding representation, is a sample of uninteracted negative examples in the average item embedding vector representation.

[0094] S23: Optimizing the embedding vector representation by optimizing the interaction loss to obtain an optimized embedding vector representation.

[0095] Each sample contains an interacting positive example and a non-interacting negative example. After multiple graph convolution learning, the adaptive moment estimation optimizer is used to update the parameters of the vector representation. After multiple rounds of optimization of the interaction loss, the vector representation changes gradually converge and the optimized embedded vector representation is obtained.

[0096] Step S3: Calculate the optimized embedding vector representation to generate the user's initial preference score for the project, rank the initial preference scores, and obtain the initial user's preference ranking for the project;

[0097] Reference Figure 2 , based on step S3, vector representations of all users and items in the candidate item pool are obtained, and the similarity between the user vector and the item vector is calculated as the user's preference ranking for the item.

[0098] Step S31: Calculate the user's initial preference score for the item based on the final vector representation of the user item learned in S2 .in, . Initial preference score The calculation is as follows:

[0099]

[0100] Initial preference scores The larger the value, the higher the user's preference for the item.

[0101] Step S32: Since the total number of items in the item candidate pool is large, in order to improve the efficiency of the algorithm, for the top-k tasks, firstly select the top-k tasks according to the initial preference scores. Keep for each user M The items with the largest preference scores will get a smaller preference score matrix and the retained items matrix .in, , , , is the number of rankings. It is worth noting that the preference score matrix The scores in the table represent the percentage of each user's reservations. M scores for each item, rather than for the same M The score of each item.

[0102] Step S33: Through the preference score matrix , calculate the user's preference ranking for the item.

[0103] Due to the influence of popularity bias, different users have different preference scores for items in terms of overall values. For example, each user has a different preference score for their favorite item, and the numerical difference is obvious. In order to eliminate the influence of the numerical value of the preference score, the ranking of the item in the user's preference score is used to measure the user's preference ranking for the item. .in, , is a set of integers, calculated as follows:

[0104]

[0105] in, Indicates the initial user About Project The preference ranking of Indicated in Project The reverse ranking of Indicates user Reserved The calculation formula shows that for each user , sort the items in reverse order according to the preference score, and the position of the item is the user The preference ranking of the item. The smaller the preference ranking value, the greater the user's preference for the item.

[0106] Step S4 re-ranks the items based on the initial preference scores and the preference rankings to obtain the recommended rankings of the items for the users:

[0107] Reference Figure 2 , according to the preference scores of different users for the items in the item pool, the relative preferences of users for different items are calculated from the perspective of the items. For each item, the relative preferences of different users for the item are ranked to obtain the recommendation tendency score of the item for the user.

[0108] Calculate the recommendation tendency ranking of the initial project to the user through the user's preference ranking of the project;

[0109] Specifically, based on the initial user's preference ranking for items in step S3, the recommendation tendency ranking of items for users is calculated from the perspective of the items. The calculation method is as follows:

[0110]

[0111] in, Display items For users The recommended tendency ranking, Indicated in In, users In ascending order, express Project The set of users with preference scores, Represents the set of all items retained in step S32.

[0112]

[0113] in, represents the norm of the matrix, Indicates taking the maximum value, Indicates taking the minimum value, Represents the preference score of all users who have interacted with the project. The calculation formula shows that for each project, users are sorted in ascending order according to their preference ranking for the project, and the position of the user is the recommendation tendency ranking of the project to the user. The smaller the recommendation tendency ranking value, the greater the recommendation tendency of the project to the user. When the user's preference ranking for the project is the same, the higher the user's preference score for the project, the greater the recommendation tendency of the project to the user, which can better distinguish the user's preference for the project.

[0114] Step S5: re-rank the user's recommendation tendency ranking of the project to obtain the user's preference ranking for the project and obtain a recommendation list:

[0115] Based on step S4, the user's preference ranking for the project is recalculated according to the recommendation tendency score of the project to the user. Specifically, based on the recommendation tendency ranking of the project to the user in step S4, the user's preference ranking for the project is converted and recalculated, and the calculation method is as follows:

[0116]

[0117] in, Indicates user About Project The calculation formula shows that for each user , sort the items in ascending order according to the recommendation tendency ranking of the items to the user, and the position of the items is the user About Project The smaller the preference ranking value, the greater the user's preference for the project. When the recommendation tendency ranking of projects to users is the same, the more users interact with the project, the smaller the user's ranking percentage in the project, and the greater the user's preference for the project, which can better distinguish the preference of the project to the user.

[0118] Take any user to be predicted as input and output the rearranged preference ranking. Figure 2 , according to the recalculated preference ranking Sort from high to low and take the first k to get the user recommended item list .in, To further utilize the project perspective, refer to Figure 2 according to Also generate a list of recommended items .in, .right and Perform intersection operations to obtain a list of recommended items that both the user and the item prefer. .in, For the recommended project list There may be some users who recommend fewer than k items. The original list of recommended items generated can be filled.

[0119] The embodiment of the present invention not only takes into account the user's preference for items, but also considers the item's recommendation preference for the user from the item's perspective, thus providing a new direction for the research on the fairness of the recommendation system.

[0120] The method proposed in the present invention has a simple flow, and only requires the user's preference score for the project to use the method of the present invention. It can be directly used as an interface in most of the current recommendation systems, and has extremely high versatility and scalability. In addition, due to its simple and efficient framework design, it hardly generates additional time consumption based on the results generated by the original recommendation system, and has extremely high operating efficiency. Compared with the prior art in the field, the present invention takes into account the accuracy and fairness of the results generated by the recommendation system from a new perspective, and has an absolute advantage in efficiency compared to other similar methods, and has excellent performance in accuracy, fairness, efficiency and other performances.

[0121] The present invention proposes a new fairness benefit indicator called "Valid Coverage@k" (abbreviated as VC@k), which is calculated as follows:

[0122]

[0123] in, For users u The generated top-k recommendation list, It is the set of items that users actually interact with in the test set.

[0124] The following is a validation of the effectiveness of the candidate item pool fairness re-ranking recommendation method ItemRec based on the project perspective of the present invention:

[0125] The present invention uses three widely used real-world interaction datasets for experiments. Table 1 shows the description information of the datasets used in this experiment.

[0126] Table 1 Experimental dataset information

[0127]

[0128] The experiment is divided into four aspects:

[0129] 1. The effect of applying the present invention to the existing recommendation method.

[0130] 2. The additional time running time overhead of the present invention over the existing recommended methods.

[0131] 3. Performance of similar fairness rearrangement methods in terms of fairness and accuracy.

[0132] 4. Time performance compared with similar efficient fairness reordering methods.

[0133] In the experiment, the embodiment of the present invention adopts two commonly used accuracy-related indicators: Recall and NDCG (Normalized Discounted Cumulative Gain), two commonly used fairness-related indicators: Coverage and Gini (Gini coefficient), and two new fairness indicators proposed by the present invention: VC@k and NV (NV is the weighted indicator of VC@k and NDCG), and NV-improv. is the growth rate of NV (unit: %). Among them, LightGCN (Light Graph Convolution Network), SGL (Self-supervised GraphLearning), NCL (Neighborhood-enriched Contrastive Learning), ENMF (EfficientNeural Matrix Factorization), SimpleX, and RecVAE in Tables 2 to 4 are all other common ranking methods. ItemRec-50-5 is a recommendation list generated from the user perspective and the project perspective using the ItemRec method proposed in the present invention, and 50 and 5 items are retained and then rearranged; ItemRec-100-5 is a recommendation list generated from the user perspective and the item perspective using the ItemRec method proposed in the present invention, and 100 and 5 items are retained and then rearranged.

[0134] Table 2 Comparison of fairness and accuracy results of dataset 1

[0135]

[0136] Table 3 Comparison of fairness and accuracy results of dataset 2

[0137]

[0138] Table 4 Comparison of fairness and accuracy results of dataset 3

[0139]

[0140] Tables 2 to 4 show the performance of existing recommendation models and methods on various data sets after using the present invention. Experimental results show that the ItemRec fairness rearrangement recommendation method proposed in the present invention greatly promotes the fairness of the recommendation results while almost not losing or even improving the accuracy of the recommendation results, and achieves better recommendation results.

[0141] Figure 3 (a)-(f) and Figure 4(a)-(f) show the performance comparison of the present invention and other fairness re-ranking methods on various data sets. The experimental results show that the present invention can achieve a better balance between accuracy and fairness indicators than other methods. Among them, FairRec (Fairness-aware Recommendation), CPFair (Consumer and Producer Fairness), P-MMF (Provider Max-min Fairness), DNaIR (Diversity-Novelty-aware Interactive Recommendation), Tax-Rank (A Taxation Perspective for FairRe-ranking), FORGE (Fairness Optimization for Ranking via Greedy Exploration) are other fairness re-ranking recommendation methods, NDCG@10 is the normalized discounted cumulative gain when the number of rankings is 10, and Gini@10 is the Gini coefficient when the number of rankings is 10.

[0142] from Figure 3 From (a) to (f), we can see that, based on the experiments on Dataset 1, the VC@10 of the ItemRec method used in the embodiment of the present invention is higher than that of other fairness re-ranking recommendation methods, which means that the method used in the embodiment of the present invention has greatly improved the sorting fairness of the corresponding basic sorting method. At the same time, the NDCG@10 and Gini@10 of the embodiment of the present invention have not increased significantly and have even decreased, which means that the method used in the embodiment of the present invention has not lost the accuracy of sorting in exchange for sorting fairness.

[0143] from Figure 4 From (a) to (f), we can see that, based on the experiments on Dataset 3, the VC@10 used in the embodiment of the present invention is higher than other fairness re-ranking recommendation methods, which means that the method used in the embodiment of the present invention has greatly improved the sorting fairness of the corresponding basic sorting method. At the same time, the NDCG@10 and Gini@10 of the embodiment of the present invention have not increased significantly and have even decreased, which means that the method used in the embodiment of the present invention has not lost the accuracy of sorting in exchange for sorting fairness.

[0144] Table 5 Results of the additional time required for rearrangement of the present invention

[0145]

[0146] Table 5 shows the additional time overhead spent by the present invention on the existing recommendation model. The experimental results show that the present invention only needs very little time overhead to achieve better recommendation results.

[0147] Table 6 Comparison table between the present invention and the existing fairness rearrangement method

[0148]

[0149] Table 6 shows the comparison of the operating efficiency between the present invention and the most efficient existing fairness rearrangement method. The results show that the present invention has excellent operating efficiency while achieving the best fairness and accuracy results.

[0150] The embodiment of the present invention also provides a candidate project pool fairness recommendation system based on project perspective, such as Figure 5 Shown include:

[0151] Acquisition module 101: Processing the interaction data of users and projects to obtain a data set for training; processing the data set to obtain a user-project interaction bipartite graph; constructing the number of users and projects to obtain a user initial embedding vector and a project initial embedding vector;

[0152] Calculation module 102: Obtain an optimized embedding vector representation through the user-item interaction bipartite graph, the user's initial embedding vector and the item's initial embedding vector; Calculate the optimized embedding vector representation to generate the user's initial preference score for the item;

[0153] Sorting module 103: rank the initial preference scores to obtain the initial user's preference ranking for the project;

[0154] Re-ranking is performed based on the initial preference scores and the initial user's preference rankings for items to obtain a recommendation tendency ranking of items for users;

[0155] The user's recommendation tendency ranking of the items is re-ranked to obtain the user's preference ranking for the items and obtain a recommendation list.

[0156] The benefit of doing this is that we can get the degree to which a project is recommended to users from a project perspective.

[0157] Compared with existing rearrangement technologies, the system described in the embodiments of the present invention provides a new sorting perspective, has better performance in sorting accuracy and fairness results than similar technologies, and has great advantages in operating efficiency. This method is superior to other existing similar methods in recommendation performance.

[0158] Through the work of the above modules, the fairness of the recommendation results is greatly promoted, and better recommendation results are achieved. Only very little time is required to achieve better recommendation results, and it also has excellent operating efficiency.

[0159] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A fair re-ranking recommendation method for a candidate project pool based on project perspective, characterized in that: The following steps are involved: S1: Process the interaction data between users and projects to obtain a data set for training; Processing the training data set to obtain a user-item interaction bipartite graph; The number of users and items is constructed to obtain the user initial embedding vector and the item initial embedding vector, including the following steps: S11: acquiring interaction data between users and items from a public data set, taking items in which users and items interact in the interaction data between users and items as positive examples, and randomly sampling items in which users and items do not interact in the interaction data between users and items to obtain negative examples, so that the number of negative examples is the same as the number of positive examples, and the positive examples and negative examples constitute a data set for training; S12: Convert the training data set into a matrix to obtain a user-item interaction bipartite graph, which is represented by a matrix A, ,in, is the set of all users to be recommended. is the candidate pool project set, is the set of real numbers; S13: Vectorize the number of users and items in the interaction data of users and items to obtain the user's initial embedding vector and the item initial embedding vector ,in, , d Indicates the dimension size of each embedding vector; S2: Obtaining an optimized embedding vector representation through the user-item interaction bipartite graph, the user initial embedding vector and the item initial embedding vector, including the following steps: S21: Through the user-item interaction bipartite graph, the user initial embedding vector and the item initial embedding vector, the graph convolution technology is used to learn the user initial embedding vector and the item initial embedding vector. The calculation formula is: in, For users to be recommended, For candidate projects, is the ordinal number of graph convolutions, represents the matrix transpose, is the user embedding vector representation without graph convolution, Embedding vector representation for items without graph convolution, Indicates The user embedding vector representation obtained by the graph convolution is Indicates The item embedding vector representation obtained by the graph convolution is For the matrix A The Laplace normalized matrix of is the diagonal matrix of the node degrees of the user graph, , is the diagonal matrix of node degrees of user and project graphs, ; S22: Calculate and optimize the interaction loss by using the user embedding vector representation, the item embedding vector representation and the interaction information of the training dataset after graph convolution learning , in, is the total number of graph convolution operations, is the average user embedding vector representation, is the average item embedding vector representation, represents the dataset used for training, Is a collection The variance of is the negative example subscript, is the sample represented by the average user embedding vector, is the sample of interactive positive examples in the average vector embedding representation, is a sample of uninteracted negative examples in the average item embedding vector representation; S23: Perform multiple graph convolution learning on the user embedding vector representation and the item embedding vector representation, use an adaptive moment estimation optimizer to update the parameters of the user embedding vector representation and the item embedding vector representation, and perform multiple rounds of optimization on the interaction loss until the user embedding vector representation and the item embedding vector representation gradually converge to obtain the optimized embedding vector representation; S3: Calculating the optimized embedding vector representation to generate the user's initial preference score for the project, ranking the initial preference scores, and obtaining the initial user's preference ranking for the project, including: S31: Calculate the user's initial preference score for the item through the optimized embedding vector representation ,in, ; S32: Reduce the initial preference scores to obtain a preference score matrix and the retained items matrix ,in, , ,in, M To retain the number, the number of retained numbers must be greater than or equal to the number of sorted numbers; S33: Sorting the preference score matrix to obtain the initial user's preference ranking for the project. in, Indicates the initial user About Project The preference ranking of Indicated in Project The reverse ranking of Indicates user Reserved M A collection of items; S4: re-ranking is performed based on the initial preference scores and the initial user's preference rankings for the items to obtain a recommendation tendency ranking of the items for the users, including: The recommendation tendency ranking of the project to the user is obtained by re-ranking the initial preference score and the initial user's preference ranking of the project. in, Display items i For users u The recommended tendency ranking, Indicated in In, users In ascending order, express Project i The set of users with preference scores, represents the norm of the matrix; S5: re-ranking the user's recommendation tendency ranking of the items to obtain the user's preference ranking for the items and obtain a recommendation list.

2. According to the method of claim 1, wherein the method comprises: Step S5 includes: adding the number of users who interacted with the project in the training dataset to the recommendation tendency ranking of the project to the user, and obtaining the user's preference ranking for the project. in, Indicates user About Project The preference ranking of , indicating that Project 's ascending ranking.

3. According to the method of claim 1, wherein the method comprises: When the number of items in the recommendation list is less than the number required for sorting, the user's initial preference ranking for the items is used. Select the most preferred item that is not recommended to fill in the gap.

4. A system for recommending fairness rearrangement of a candidate item pool based on a project perspective, for executing the method for recommending fairness rearrangement of a candidate item pool based on a project perspective as claimed in any one of claims 1 to 3, characterized in that: include: Acquisition module: processes the interaction data between users and projects to obtain data sets for training; Processing the data set to obtain a user-item interaction bipartite graph; constructing the number of users and items to obtain a user initial embedding vector and an item initial embedding vector; The interaction data between users and items is obtained from a public data set, and the items in which users and items interact in the interaction data between users and items are taken as positive examples. Negative examples are randomly sampled from the items in which users and items do not interact in the interaction data between users and items, so that the number of negative examples is the same as the number of positive examples. The positive examples and negative examples constitute a data set for training; the data set for training is converted into a matrix to obtain a user-item interaction bipartite graph, and the user-item interaction bipartite graph is represented as a matrix A, ,in, is the set of all users to be recommended. is the candidate pool project set, is a real number set; the number of users and items in the interaction data of users and items is vectorized to obtain the user's initial embedding vector and the item initial embedding vector ,in, , d Indicates the dimension size of each embedding vector; Calculation module: Obtain optimized embedding vector representation through the user-item interaction bipartite graph, the user initial embedding vector and the project initial embedding vector; Use graph convolution technology to learn the user initial embedding vector and the project initial embedding vector through the user-item interaction bipartite graph, the user initial embedding vector and the project initial embedding vector, and the calculation formula is: in, For users to be recommended, For candidate projects, is the ordinal number of graph convolutions, represents the matrix transpose, is the user embedding vector representation without graph convolution, Embedding vector representation for items without graph convolution, Indicates The user embedding vector representation obtained by the graph convolution is Indicates The item embedding vector representation obtained by the graph convolution is For the matrix A The Laplace normalized matrix of is the diagonal matrix of the node degrees of the user graph, , is the diagonal matrix of node degrees of user and project graphs, ; Calculate and optimize the interaction loss by using the user embedding vector representation, item embedding vector representation and the interaction information of the training dataset after graph convolution learning , in, is the total number of graph convolution operations, is the average user embedding vector representation, is the average item embedding vector representation, represents the dataset used for training, Is a collection The variance of is the negative example subscript, is the sample represented by the average user embedding vector, is the sample of interactive positive examples in the average vector embedding representation, The method comprises the following steps: performing multiple graph convolution learning on the user embedding vector representation and the project embedding vector representation, updating the parameters of the user embedding vector representation and the project embedding vector representation by using an adaptive moment estimation optimizer, and performing multiple rounds of optimization on the interaction loss until the user embedding vector representation and the project embedding vector representation gradually converge to obtain an optimized embedding vector representation; calculating the optimized embedding vector representation to generate the user's initial preference score for the project; and calculating the user's initial preference score for the project by using the optimized embedding vector representation. ,in, ; Sorting module: rank the initial preference scores to obtain the initial user preference ranking for the project; reduce the initial preference scores to obtain the preference score matrix and the retained items matrix ,in, , ,in, M The number of items to be retained is greater than or equal to the number of items to be sorted. The preference score matrix is ​​used to sort the items to obtain the initial user's preference ranking for the items. in, Indicates the initial user About Project The preference ranking of Indicated in Project The reverse ranking of Indicates user Reserved M A collection of items; Re-ranking is performed based on the initial preference scores and the initial user's preference rankings for the items to obtain a recommendation tendency ranking of the items for the users; re-ranking is performed based on the initial preference scores and the initial user's preference rankings for the items to obtain a recommendation tendency ranking of the items for the users, in, Display items i For users u The recommended tendency ranking, Indicated in In, users In ascending order, express Project i The set of users with preference scores, represents the norm of the matrix; The user's recommendation tendency ranking of the items is re-ranked to obtain the user's preference ranking for the items and obtain a recommendation list.

Images