A two-sided fairness approach for large language model recommendation systems

Abstract

This invention discloses a bilateral fairness method for large language model recommendation systems, which relates to the field of machine learning technology. The method comprises the following steps: collecting user behavior data and project data; processing the user behavior data and project data to generate preliminary embeddings for users, preliminary embeddings for projects, and similar user embeddings, respectively; generating enhanced embeddings for users and projects using a large language model; fusing the preliminary embeddings with the enhanced embeddings to obtain fused user embeddings and fused project embeddings; performing a dot product operation on the fused user embeddings and fused project embeddings to obtain a recommendation score, introducing an interaction count penalty factor to process the recommendation score to obtain a weighted recommendation score; and processing the weighted recommendation score using a bilateral fairness optimization strategy to generate a final recommendation list. This method can accurately identify and address unfairness.

Description

Technical Field

[0001] The present invention relates to the field of machine learning technology, and in particular to a bilateral fairness method for large language model recommendation system scenarios. Background Art

[0002] In recent years, with the rapid spread of the internet, the amount of data has been increasing rapidly. Faced with the intractable challenge of processing this massive amount of data, recommender systems have emerged. Recommender systems permeate nearly every aspect of our lives, from e-commerce and multimedia to educational course selection. Recommender systems estimate users' interests and preferences based on their historical interactions and provide personalized recommendations, thereby improving both efficiency and accuracy. Although recommender systems have become one of the most successful examples of AI practice, some research has highlighted various types of fairness issues. One area of ​​research focuses on unfairness among projects, particularly popularity bias in recommendations, whereby popular projects receive more exposure than less popular projects. This issue stems primarily from the fact that during model training, popular projects receive more attention due to their greater interaction data than less popular projects, leading the model to be more biased in recommending these products. Another area of ​​research focuses on unfairness on the user side, where users are divided into active and inactive groups based on their level of activity in the system. Typically, systems perform better for active users than inactive ones. This is likely because when training a model, users who interact more frequently with the platform generate more interaction data than those who interact less frequently, allowing the model to better understand and prioritize the preferences of these users. This discriminatory treatment, such as differences in item exposure and user performance due to data heterogeneity, can lead to individual economic disparities and even amplify inherent biases in the model.

[0003] In order to better distinguish these unfairness issues, existing research works divide them into: user fairness, project fairness and bilateral fairness according to the research objects. Among them, user fairness is divided into individual fairness and group fairness. The former emphasizes that the model pays equal attention to the personalized preferences of each user, while the latter emphasizes that user groups with different attributes should receive the same treatment from the model. Most existing works use methods such as adding regularization terms, constrained optimization, and balancing data sets; project fairness is also divided into individual fairness and group fairness, which is usually reflected in the proportion of individuals and groups in the recommendation list. Most of the work is achieved by proposing new fair re-ranking algorithms and reinforcement learning; bilateral fairness refers to considering user fairness and project fairness at the same time. Most of the current work uses the addition of regularization and constraints.

[0004] With the rapid development of large language models, a growing number of studies have explored their integration with recommendation systems, leading to the emergence of large language model recommendation systems. These recommendation systems possess enhanced understanding and generation capabilities. They can not only make recommendations based on user behavioral data but also deeply explore user needs through techniques such as natural language processing and conversation modeling, resulting in more personalized recommendations. However, compared to traditional recommendation systems, fairness issues in large language model recommendation systems are more complex. Large language models can generate personalized recommendations by understanding user information such as multi-turn conversations, sentiment, and language expressions. This flexible modeling approach offers more recommendation possibilities for inactive users, as the model can infer user interests through language and context even in the absence of sufficient behavioral data. However, because large language models rely on training data, they can result in poor recommendation quality for certain user groups (such as inactive or cold-start users) or overfit certain user characteristics, leading to unfairness. Furthermore, large language models typically comprehensively analyze a user's historical data, current context, and textual descriptions of items (such as product information or movie plots) to generate personalized recommendations for each user. However, the issue of item exposure still exists. Since popular items usually have more interaction data and a broad audience base, they tend to be recommended first, which may cause unpopular items to not receive enough attention, thus limiting the diversity of the recommendation system.

[0005] While much research has been conducted on large language model recommendation systems, this work has focused primarily on improving performance, particularly in terms of recommendation accuracy, personalized recommendation quality, and system response speed. Only a few studies have addressed the fairness issues common to traditional recommendation systems in large language model recommendation systems, and these studies focus solely on user fairness or item fairness, lacking research on bilateral fairness. Therefore, designing a bilateral fairness approach that can simultaneously mitigate both user and item fairness issues in large language model recommendation systems has become a pressing issue in this field. Summary of the Invention

[0006] In order to solve the technical problems existing in the above-mentioned prior art, the present invention proposes a bilateral fairness method for the large language model recommendation system scenario, which can effectively alleviate the bilateral fairness problem of the recommendation system while maintaining the overall performance of the large language model recommendation system.

[0007] To achieve the above objectives, the present invention provides a two-sided fairness method for large language model recommendation system scenarios, including:

[0008] Collecting user behavior data and project data, wherein the user behavior data includes user interaction information on projects, and the project data includes detailed information on each project;

[0009] Processing the user behavior data and the project data to generate preliminary embeddings for users, preliminary embeddings for projects, and similar user embeddings;

[0010] Leverage large language models to generate enhanced embeddings for users and items;

[0011] Fusing the user's preliminary embedding, the similar user embedding, and the user's enhanced embedding to generate a fused user embedding, and fusing the item's preliminary embedding and the item's enhanced embedding to generate a fused item embedding;

[0012] Performing a dot product operation on the fused user embedding and the fused item embedding to obtain a recommendation score, introducing an interaction count penalty factor to process the recommendation score to obtain a weighted recommendation score;

[0013] The weighted recommendation scores are processed using a bilateral fairness optimization strategy to generate a final recommendation list.

[0014] Preferably, generating the preliminary embedding of the user and the preliminary embedding of the item comprises:

[0015] The collaborative filtering method is used to process each user behavior data and item data respectively to obtain the preliminary embedding of the user and the preliminary embedding of the item.

[0016] Preferably, generating the similar user embeddings comprises:

[0017] The similar user embedding is generated by calculating the similarity between the user behavior data, specifically:

[0018]

[0019] Where, Embedding for similar users, u i Represents user i,u j represents user j, Sim(u i ,u j ) represents user u i with u j The similarity between For user u j The initial embedding of Top-K(u i ) indicates that the user u i The K most similar users.

[0020] Preferably, a weighted average method or a splicing method is used to generate the fused user embedding and the fused item embedding;

[0021] Wherein, the weighted average method is:

[0022]

[0023] Where α, β, γ and α ′ , β ′ are all hyperparameters, To integrate user embedding, For user u i The initial embedding of Embedding for similar users, For user u i Enhanced embedding of Embedded for fusion projects, For project i j The initial embedding of For project i j Enhanced embedding of

[0024] The splicing method is:

[0025]

[0026] Where concat means concatenating multiple embedding vectors into a larger embedding.

[0027] Preferably, the interaction number penalty factor is:

[0028]

[0029] Where c j Indicates project i j The number of times it is interacted by all users, c min and c max Represents the minimum and maximum number of interactions of all items, P j is the interaction penalty factor, and δ is the scaling factor.

[0030] Preferably, obtaining the weighted recommendation score includes:

[0031] y i ′ j =y ij ·P j ;

[0032] Where y ij Represents user u i For Project i j The original recommendation score, y i ′ j is the weighted recommendation score.

[0033] Preferably, the bilateral fairness optimization strategy includes:

[0034] Quantify the user unfairness index and the project exposure unfairness index, comprehensively consider the recommendation score, user fairness and project exposure fairness, and obtain a multi-objective optimization function;

[0035] Wherein, the multi-objective optimization function is:

[0036]

[0037] In the formula, λ1 and λ2 are adjustment coefficients used to balance the relationship between score maximization and fairness optimization. is a multi-objective optimization function, ΔNDCG is the performance index difference between the active user group and the inactive user group, ΔExposure is the exposure difference between popular items and unpopular items in the recommendation list, y i ′ j is the weighted recommendation score, N is the number of users, and M is the number of items.

[0038] Preferably, quantifying the user unfairness indicator includes:

[0039] ΔNDCG=|NDCG(u active )-NDCG(u inactive )|;

[0040] Where, NDCG(u active ) and NDCG(u inactive ) represent the NDCG values ​​of active users and inactive users respectively;

[0041] Quantifying indicators of unfair project exposure includes:

[0042] ΔExposure=|Exposure popular -Exposure cold |;

[0043] Where, Exposure popular and Exposure cold Respectively represent the proportion of popular projects and unpopular projects in the recommendation list.

[0044] Preferably, the final recommendation list is:

[0045]

[0046] Where, For user u i The final recommendation list contains the items after fairness optimization; Rank() is the ranking function.

[0047] Compared with the prior art, the present invention has the following advantages and technical effects:

[0048] (1) This invention effectively alleviates the bias problem in the recommendation system by introducing a bilateral optimization strategy for user fairness and project fairness. Compared with traditional methods, this method can accurately identify and adjust unfairness, ensuring that different user groups and project categories can participate fairly in the recommendation process. It achieves a significant balance between diversity and fairness, improves users' trust and satisfaction with the system, and contributes to the long-term healthy development of the platform;

[0049] (2) This paper combines collaborative filtering embedding and large language model embedding to propose a more expressive embedding generation method that organically integrates user behavior with project text information. By adding similar user embeddings, the user's potential preferences are further explored, making the recommendation results more diverse and accurate. In addition, during the recommendation list generation process, a re-ranking algorithm with the number of interactions as a penalty factor is designed to reduce excessive bias towards popular projects, thereby increasing the exposure opportunities of unpopular projects, and enhancing the diversity of recommendations while ensuring the quality of recommendations;

[0050] (3) This invention adopts a multi-objective optimization method, combining recommendation score maximization with bilateral fairness optimization, and flexibly adapting to different scenario requirements by adjusting hyperparameter weights. This method not only dynamically balances recommendation quality and fairness, but also has strong scalability and is suitable for the deployment of various recommendation systems. Through this optimization mechanism, the platform can ensure that the recommendation system has higher robustness and fairness while meeting the personalized needs of users, thereby improving the overall user experience and ecosystem vitality. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0052] Figure 1 This is a flowchart of a bilateral fairness method for a large language model recommendation system scenario according to an embodiment of the present invention;

[0053] Figure 2 This is a framework diagram of a large language model recommendation system that alleviates the bilateral fairness issue according to an embodiment of the present invention;

[0054] Figure 3 Schematic diagram of improved embedded fusion enhancement according to an embodiment of the present invention;

[0055] Figure 4 This is an improved multi-objective optimization flowchart of an embodiment of the present invention. DETAILED DESCRIPTION

[0056] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0057] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0058] This embodiment proposes a bilateral fairness method for large language model recommendation system scenarios, such as Figure 1-Figure 2 ,include:

[0059] Collect user behavior data and project data, where user behavior data includes user interaction information on projects, and project data includes detailed information about each project;

[0060] Perform data processing on user behavior data and project data to generate preliminary embeddings for users, preliminary embeddings for projects, and similar user embeddings respectively;

[0061] Leverage large language models to generate enhanced embeddings for users and items;

[0062] Fusing the user's preliminary embedding, similar user embeddings, and the user's enhanced embedding to generate a fused user embedding; fusing the item's preliminary embedding with the item's enhanced embedding to generate a fused item embedding;

[0063] Perform a dot product operation on the fused user embedding and the fused item embedding to obtain a recommendation score. Then, introduce an interaction penalty factor to process the recommendation score to obtain a weighted recommendation score.

[0064] The weighted recommendation scores are processed through the bilateral fairness optimization strategy to generate the final recommendation list.

[0065] This embodiment first collects user behavior data and project data and generates preliminary embeddings for users and projects. Secondly, similar user embeddings are generated based on similarity calculations. Subsequently, a large language model is used to generate enhanced embeddings for users and projects, which are then fused with the embeddings generated by collaborative filtering. Next, an interaction penalty factor is introduced to reduce over-recommendations for popular projects. Then, a bilateral fairness optimization strategy is implemented to balance user fairness and project exposure fairness. Finally, a final recommendation list is generated to improve the accuracy, diversity, and fairness of the recommendations, ensuring that the fairness requirements of users and projects are met while maximizing the recommendation score.

[0066] Specifically, in this example recommendation system, we first need to collect user behavior data and project information. User behavior data includes user interaction information on projects, including clicks, ratings, etc. Project data includes detailed information about each project, such as title, description, category, etc.

[0067] Assuming there are N users and M items, user behavior data can be expressed as:

[0068]

[0069] Where u i represents user i, i j Represents project j, r ij Represents user u i For Project i j Feedback value, D user In this embodiment, the user behavior dataset is used to train and verify the recommendation model.

[0070] Furthermore, generating preliminary embeddings of users and items includes:

[0071] The collaborative filtering method is used to process each user behavior data and item data respectively to obtain the preliminary embedding of the user and the preliminary embedding of the item.

[0072] Specifically, assuming that the embedding of each user and item is obtained through a matrix decomposition algorithm, the preliminary embedding of the user is generated through the interaction data between the user and the item and initial project embedding These preliminary embeddings are the basis of the recommendation system and reflect the interaction history characteristics of users and items.

[0073] Furthermore, similar user embeddings are generated, including:

[0074] Calculate the similarity between the user behavior data and generate the similar user embedding, specifically:

[0075]

[0076] Where, Embedding for similar users, u i Represents user i,u j represents user j, Sim(u i ,u j ) represents user u i with u j The similarity between For user u j The initial embedding of Top-K(u i ) indicates that the user u iThe K most similar users.

[0077] Specifically, the cosine similarity calculation method is used in this embodiment:

[0078]

[0079] Among them, Sim(u i ,u j ) represents user u i with u j The similarity between For user u i Initial embedding.

[0080] Further, if Figure 3 , using the large language model to generate enhanced embeddings for users and items. The large language model generates richer semantic embeddings by parsing the user's behavior history and the text information of the item. For example, for user u i , user behavior sequences, such as clicked items, rated content, etc., can be input into the large language model to obtain the user's enhanced embedding Similarly, the text description of the item, such as title, category, introduction, etc., can be input into the large language model to obtain the enhanced embedding of the item

[0081] Fuse the user's preliminary embedding, similar user embedding, and user's enhanced embedding to generate a fused user embedding; fuse the item's preliminary embedding and item's enhanced embedding to generate a fused item embedding;

[0082] To fully utilize collaborative filtering and embeddings generated by a large language model, this embodiment combines them by weighted averaging or concatenation:

[0083] Among them, the weighted average method is:

[0084]

[0085] Where α, β, γ and α ′ , β ′ are all hyperparameters, To integrate user embedding, For user u i The initial embedding of Embedding for similar users, For user u i Enhanced embedding of Embedded for fusion projects, For project i j The initial embedding of For project i j Enhanced embedding of

[0086] The splicing method is:

[0087]

[0088] Where concat means concatenating multiple embedding vectors into a larger embedding.

[0089] This example combines collaborative filtering embeddings with large language model embeddings to propose a more expressive embedding generation method, organically integrating user behavior with item text information. By adding similar user embeddings, it further explores users' potential preferences, making recommendation results more diverse and accurate. Furthermore, during the recommendation list generation process, a re-ranking algorithm using the number of interactions as a penalty factor is designed to reduce excessive bias towards popular items, thereby increasing exposure for less popular items, enhancing recommendation diversity while maintaining quality.

[0090] Furthermore, in order to reduce over-recommendation of popular items, this embodiment introduces an interaction penalty factor. The interaction penalty factor is:

[0091]

[0092] Where c j Indicates project i j The number of times it is interacted by all users, c min and c max Represents the minimum and maximum number of interactions of all items, P j is the penalty factor for the number of interactions, and δ is the scaling factor, which is used to control the impact of the number of interactions on the penalty factor.

[0093] Furthermore, obtaining the weighted recommendation score includes:

[0094] y i ′ j =y ij ·P j ;

[0095] Where y ij Represents user u i For Project i j The original recommendation score, y i ′ j is the weighted recommendation score.

[0096] Specifically, when generating the recommendation score, the penalty factor is added to the original score, thereby reducing the recommendation score for popular items and increasing the recommendation score for unpopular items, thereby improving the ranking of unpopular items in the recommendation list and increasing the recommendation exposure of unpopular items.

[0097] Furthermore, if Figure 4 , which is the improved multi-objective optimization flowchart of this embodiment. In order to ensure the overall performance of the large language model recommendation system while alleviating unfairness, this embodiment introduces a bilateral fairness optimization strategy, which optimizes user fairness and project exposure fairness while maximizing the recommendation score.

[0098] Two-sided fairness optimization strategies include:

[0099] Quantify the user unfairness index and the project exposure unfairness index, comprehensively consider the recommendation score, user fairness and project exposure fairness, and obtain a multi-objective optimization function;

[0100] Among them, the multi-objective optimization function is:

[0101]

[0102] In the formula, λ1 and λ2 are adjustment coefficients used to balance the relationship between score maximization and fairness optimization. is a multi-objective optimization function, ΔNDCG is the performance index difference between the active user group and the inactive user group, ΔExposure is the exposure difference between popular items and unpopular items in the recommendation list, y i ′ j is the weighted recommendation score, N is the number of users, and M is the number of items.

[0103] Specifically, the quantification of user unfairness indicators includes:

[0104] ΔNDCG=|NDCG(u active )-NDCG(u inactive )|;

[0105] Where, NDCG(u active ) and NDCG(u inactive ) represent the NDCG values ​​of active users and inactive users respectively;

[0106] Quantifying indicators of unfair project exposure includes:

[0107] ΔExposure=|Exposure popular -Exposure cold |;

[0108] Where, Exposure popular and Exposure cold Respectively represent the proportion of popular projects and unpopular projects in the recommendation list.

[0109] Furthermore, after bilateral fairness optimization, the final recommendation list is generated by sorting:

[0110]

[0111] Where, For user u i The final recommendation list contains the items after fairness optimization; Rank() is the sorting function, which is based on the recommendation score y i ′ j Sort.

[0112] This example effectively mitigates bias in the recommendation system by introducing a bilateral optimization strategy for both user fairness and project fairness. Compared to traditional methods, this example method accurately identifies and addresses unfairness, ensuring that diverse user groups and project categories can participate fairly in the recommendation process. This achieves a significant balance between diversity and fairness, enhancing user trust and satisfaction with the system and contributing to the long-term healthy development of the platform.

[0113] This embodiment employs a multi-objective optimization approach, combining recommendation score maximization with bilateral fairness optimization. By adjusting hyperparameter weights, it flexibly adapts to the needs of different scenarios. This approach not only dynamically balances recommendation quality and fairness but also offers strong scalability, making it suitable for deployment in various recommendation systems. Through this optimization mechanism, the platform can meet user personalized needs while ensuring a more robust and fair recommendation system, thereby enhancing the overall user experience and ecosystem vitality.

[0114] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A bilateral fairness method for large language model recommendation system scenarios, characterized by: include: Collecting user behavior data and project data, wherein the user behavior data includes user interaction information on projects, and the project data includes detailed information on each project; Processing the user behavior data and the project data to generate preliminary embeddings for users, preliminary embeddings for projects, and similar user embeddings; Leverage large language models to generate enhanced embeddings for users and items; Fusing the user's preliminary embedding, the similar user embedding, and the user's enhanced embedding to generate a fused user embedding, and fusing the item's preliminary embedding and the item's enhanced embedding to generate a fused item embedding; Performing a dot product operation on the fused user embedding and the fused item embedding to obtain a recommendation score, introducing an interaction count penalty factor to process the recommendation score to obtain a weighted recommendation score; Processing the weighted recommendation scores through a bilateral fairness optimization strategy to generate a final recommendation list; The interaction penalty factor is: ； Where, Display items The number of times it is interacted with by all users, and Respectively represent the minimum and maximum number of interactions for all items, is the interaction penalty factor, is the scaling factor; The bilateral fairness optimization strategy includes: Quantify the user unfairness index and the project exposure unfairness index, comprehensively consider the recommendation score, user fairness and project exposure fairness, and obtain a multi-objective optimization function; Wherein, the multi-objective optimization function is: ； Where, and Both are adjustment coefficients used to balance the relationship between score maximization and fairness optimization. is a multi-objective optimization function, is the performance indicator difference between the active user group and the inactive user group, The difference in exposure between popular and unpopular projects in the recommendation list. is the weighted recommendation score, is the number of users, is the number of items.

2. The two-sided fairness method for large language model recommendation system scenarios according to claim 1, characterized in that: Generating a preliminary embedding of the user and a preliminary embedding of the item, including: The collaborative filtering method is used to process each user behavior data and item data respectively to obtain the preliminary embedding of the user and the preliminary embedding of the item.

3. The two-sided fairness method for large language model recommendation system scenarios according to claim 1, characterized in that: Generating the similar user embedding includes: The similar user embedding is generated by calculating the similarity between the user behavior data, specifically: ； Where, Embedding for similar users, Represents a user , Represents a user , Represents a user and The similarity between For users The initial embedding of Represents user Most similar users.

4. The two-sided fairness method for large language model recommendation system scenarios according to claim 3, characterized in that: Generating the fused user embedding and the fused item embedding by adopting a weighted average method or a splicing method; The weighted average method is: ； ； Where, , , and , are all hyperparameters, To integrate user embedding, For users The initial embedding of Embedding for similar users, For users Enhanced embedding of Embedded for fusion projects, For the project The initial embedding of For the project Enhanced embedding of The splicing method is: ； ； Where, It represents concatenating multiple embedding vectors into a larger embedding.

5. The two-sided fairness method for large language model recommendation system scenarios according to claim 1, characterized in that: Obtaining the weighted recommendation score includes: ； Where, Represents a user About the project The original recommendation score of is the weighted recommendation score.

6. The two-sided fairness method for large language model recommendation system scenarios according to claim 1, characterized in that: Quantifying user unfairness indicators includes: ； Where, and Represents active users and inactive users respectively value; Quantifying indicators of unfair project exposure includes: ； Where, and Respectively represent the proportion of popular projects and unpopular projects in the recommendation list.

7. The two-sided fairness method for large language model recommendation system scenarios according to claim 6, characterized in that: The final recommendation list is: ； Where, For users The final recommendation list includes fairness-optimized projects; is the sorting function.

Images