Fair recommendation method and system based on large model self-gaming

Abstract

The invention discloses a fair recommendation method and system based on large model self-gaming, and the method comprises the steps: a self-gaming data generation link: building a preference data set based on user historical behavior data as a positive sample and a recommendation result generated by combining a current model recommendation strategy as a negative sample, samples with too high similarity between positive and negative samples are removed through data filtering; and a supervised training link: on the basis of a pre-trained large language model, performing supervised training by using the preference data set, optimizing model parameters through cross entropy loss, and generating a preliminary recommendation strategy. The invention aims to optimize the long tail phenomenon of the large language model in the recommendation task by introducing a self-game mechanism, and improve the diversity and fairness of the recommendation system. Through the self-generation training data of the model, the performance of the recommendation system in different user groups and application scenes is improved, the personalization and diversity of the recommendation system are enhanced, and the stability and fairness of the model are improved in long-term iteration.

Description

Technical Field

[0001] The present application relates to the technical field of recommendation systems, and more specifically to a fair recommendation method and system based on large-model self-game. Background Art

[0002] In large language model recommendation systems, over-concentration on recommending popular items has become a major obstacle that limits the fairness and diversity of recommendation systems. Traditional fine-tuning methods, such as supervised learning and direct preference optimization, are often used in the training process of large models to optimize model output by using labeled recommendation data or combining recommendation data with random negative samples. However, these methods are prone to introduce bias during the fine-tuning process, resulting in excessive recommendation of popular items, thus affecting the diversity and fairness of recommendation results. Although they can improve the recommendation ability of the model, they are highly dependent on bias and usually cannot effectively alleviate the bias problem in the recommendation. They often need to rely on external data, additional annotations, or complex post-processing steps, which increases computational overhead and system complexity.

[0003] The self-game mechanism originated in the field of reinforcement learning and was originally used for zero-sum game problems. By allowing multiple agents to compete against each other in the same environment, the agents are prompted to optimize their strategies to achieve a balance in the game. Self-games have adaptive and self-optimizing properties, making them particularly effective in solving tasks with long-term feedback or high uncertainty. Although self-games have been widely used in reinforcement learning and game theory, their potential in recommendation systems has not yet been fully explored. Most current research on recommendation systems focuses on optimization strategies based on user historical behavior or item similarity, and fails to fully consider self-optimizing recommendation strategies through self-games to improve the fairness and content diversity of the system. Summary of the invention

[0004] In this embodiment, a fair recommendation method, system, electronic device and storage medium based on large model self-game are provided to solve the problem of bias of popular items faced by current recommendation systems in related technologies.

[0005] In a first aspect, an embodiment of the present invention provides a fair recommendation method based on a large model self-game, and the fair recommendation method based on a large model self-game includes: In the self-game data generation phase, based on the user's historical behavior data as positive samples and the recommendation results generated by the current model recommendation strategy as negative samples, a preference data set is constructed, and samples with too high similarity between positive and negative samples are removed through data filtering; In the supervised training phase, based on the pre-trained large language model, the preference data set is used for supervised training, the model parameters are optimized through cross entropy loss, and a preliminary recommendation strategy is generated; In the direct preference optimization phase, based on the preference data set generated by self-game, the recommendation strategy is directly adjusted by optimizing the objective function to suppress over-recommendation of popular items and balance the long-tail distribution. The optimization objective function is constructed by comparing the output probability difference between the current model and the reference model on positive and negative samples.

[0006] In an optional embodiment, in the self-game data generation link, positive samples include user click, browse or rating behavior data; negative samples are generated through a model iteration recommendation strategy and dynamically matched with user behavior data.

[0007] In an optional embodiment, the data filtering includes: Calculate the Euclidean distance between the embedding vectors of the target recommended item and the predicted recommended item. If the distance is less than the preset threshold, the corresponding positive and negative samples are removed.

[0008] In an optional embodiment, the model initialization of the supervised training phase uses pre-trained large language model parameters, and the parameters are updated through the Adam optimizer.

[0009] In an optional embodiment, in the supervised training phase, the cross entropy loss function is defined as: ; in, is the true result distribution, Output probability distribution for prediction.

[0010] In an optional embodiment, the optimization objective function of the direct preference optimization link is: ; in, For the current model, is the reference model, is the weight parameter; is a positive sample, is a negative sample, For input context.

[0011] In an optional embodiment, the direct preference optimization step further includes generating diversified preference data through a self-game mechanism to increase the probability of recommending low-frequency items.

[0012] Compared with the prior art, the fair recommendation method based on large model self-game of the present invention has the following beneficial effects: The present invention aims to optimize the long-tail phenomenon of large language models in recommendation tasks and improve the diversity and fairness of recommendation systems by introducing a self-game mechanism. Through the feedback mechanism of self-game, this method can adaptively adjust the recommended content without the need for additional data or prior knowledge, thereby reducing the over-recommendation of popular items and improving the distribution of recommendation results. The goal of this method is to improve the performance of the recommendation system in different user groups and application scenarios through the self-generated training data of the model, enhance the personalization and diversity of the recommendation system, and improve the stability and fairness of the model in long-term iterations.

[0013] In a second aspect, an embodiment of the present invention provides a fair recommendation system based on large model self-game, including: The self-game data generation module is configured to: build a preference data set based on user historical behavior data as positive samples and the recommendation results generated by the current model recommendation strategy as negative samples, and remove samples with too high similarity between positive and negative samples through data filtering; The supervised training module is configured to: perform supervised training using the preference data set based on the pre-trained large language model, optimize model parameters through cross entropy loss, and generate a preliminary recommendation strategy; The direct preference optimization module is configured to: directly adjust the recommendation strategy by optimizing the objective function based on the preference data set generated by self-game, suppress the over-recommendation of popular items and balance the long-tail distribution, wherein the optimization objective function is constructed by comparing the output probability difference between the current model and the reference model on positive and negative samples.

[0014] In a third aspect, an embodiment of the present invention provides an electronic device, comprising a processor, a communication interface, a memory and a bus, wherein the processor, the communication interface and the memory communicate with each other through the bus, and the processor can call logic instructions in the memory to execute the steps of the method provided in the first aspect.

[0015] In a fourth aspect, an embodiment of the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the fair recommendation method based on large model self-game as described in the first aspect.

[0016] Compared with the prior art, the beneficial effects of the fair recommendation system, electronic device and storage medium based on large model self-game of the present invention are the same as the fair recommendation method based on large model self-game described in the first aspect, so they will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. 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.

[0018] Figure 1 It is a flow chart of a fair recommendation method based on large model self-game in an embodiment of the present invention; Figure 2 A schematic diagram comparing SEAR with other large model fine-tuning techniques in an embodiment of the present invention; Figure 3 It is a principle diagram of a fair recommendation method based on large model self-game in an embodiment of the present invention; Figure 4 This is a diagram showing the performance changes of the model in the embodiment of the present invention; Figure 5 It is a structural block diagram of a fair recommendation system based on large model self-game in an embodiment of the present invention; Figure 6 1 is a structural block diagram of an electronic device in an embodiment of the present invention. DETAILED DESCRIPTION

[0019] In order to more clearly understand the purpose, technical solutions and advantages of the present application, the present application is described and illustrated below in conjunction with the accompanying drawings and embodiments.

[0020] Unless otherwise defined, the technical terms or scientific terms involved in this application shall have the general meaning understood by people with general skills in the technical field to which this application belongs. The words "one", "a", "the", "these" and the like in this application do not indicate a quantitative limitation, and they may be singular or plural. The terms "include", "comprise", "have" and any variants thereof involved in this application are intended to cover non-exclusive inclusions; for example, a process, method and system, product or device comprising a series of steps or modules (units) is not limited to the listed steps or modules (units), but may include unlisted steps or modules (units), or may include other steps or modules (units) inherent to these processes, methods, products or devices. The words "connect", "connected", "coupled" and the like involved in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The "multiple" involved in this application refers to two or more. "And / or" describes the association relationship of associated objects, indicating that there may be three relationships, for example, "A and / or B" may mean: A exists alone, A and B exist at the same time, and B exists alone. Generally, the character " / " indicates that the objects associated with each other are in an "or" relationship. The terms "first", "second", "third", etc. in this application are only used to distinguish similar objects and do not represent a specific ordering of the objects.

[0021] In an embodiment of the present invention, a fair recommendation method (SEAR) based on large model self-game is provided. Figure 1 is a flow chart of a fair recommendation method based on large model self-game of the present invention, such as Figures 1 to 3 As shown, the process includes the following steps: S100, self-game data generation stage, based on the user's historical behavior data as a positive sample, combined with the recommendation results generated by the current model recommendation strategy as a negative sample, to build a preference data set, and remove samples with too high similarity between positive and negative samples through data filtering; It should be noted that in the self-game data generation process, positive samples include users' click, browsing or rating behavior data; negative samples are generated through model iteration recommendation strategies and dynamically matched with user behavior data.

[0022] Data filtering includes: Calculate the Euclidean distance between the embedding vectors of the target recommended item and the predicted recommended item. If the distance is less than the preset threshold, the corresponding positive and negative samples are removed.

[0023] Specifically, during the training process of the SEAR method, the initial recommendation strategy is usually obtained through supervised learning methods. The large language model (LLM) first accepts a set of labeled data (such as historical user behavior data), and then fine-tunes the model with this data. The supervised trained model usually predicts recommended items based on the user's historical preferences and the characteristics of the items. The SEAR method debiased the supervised training model through subsequent self-play fine-tuning.

[0024] The generation of the recommendation data set is to generate a data set with preferences by combining the model's current recommendation strategy with the user behavior in the real data set. In this process: Real data as positive samples: The user's click, browse, and rating behavior data in the historical recommendation system are regarded as positive samples. These positive samples represent the user's preference for certain items and are the targets that need to be focused on during the model learning process.

[0025] Model recommendation results as negative samples: The recommendation results generated by the model based on the initial recommendation strategy or the self-game process are considered negative samples. These items are usually recommended items that the model has not optimized or are biased, so they need to be optimized during the training process.

[0026] The preference dataset combines real user behavior data (positive samples) and the recommendation results output by the model (negative samples), providing diverse training data for the model. Training on this dataset enables the model to learn user preferences while avoiding the spread of bias and improving the diversity and fairness of recommendations.

[0027] In order to ensure the quality of the recommendation dataset, an additional data filtering step is introduced to avoid optimization problems caused by excessive similarity between positive and negative samples. Specifically, the sentence embedding model is used to encode the target recommended items and the predicted recommended items to obtain their embeddings. . Then the similarity between the two is calculated by Euclidean distance, that is: ; like If it is less than a pre-set threshold, the similarity between positive and negative samples is considered to be too high, which is not conducive to the preference optimization stage and is removed from the data set.

[0028] S200, supervised training phase, based on the pre-trained large language model, supervised training is performed using the preference data set, model parameters are optimized through cross entropy loss, and a preliminary recommendation strategy is generated; It should be noted that the model initialization in the supervised training phase uses the pre-trained large language model parameters, and the parameters are updated through the Adam optimizer.

[0029] Specifically, supervised learning refers to training a model with data with known labels so that the model can predict accurate outputs based on the input data. In a recommendation system, the task of supervised learning is to let the model learn the user's preferences based on the user's historical behavior data (such as clicks, ratings, purchase records, etc.) and generate recommendation results that match these preferences. Specifically, supervised learning optimizes the model parameters by minimizing the loss function so that the model can match the user's preferences as much as possible when recommending items. It includes the following steps: Model initialization: Based on the pre-trained large language model, the model parameters are randomly initialized. The pre-trained model has mastered a lot of general language knowledge, and the fine-tuning in this stage aims to improve the performance of the model in specific recommendation tasks, so that it can better capture user preferences and achieve accurate recommendations.

[0030] Forward propagation and loss calculation: After the user's historical behavior data is input into the model, the model generates a predicted output through forward propagation. Probability distribution of predicted output Distribution of the actual results The error between them is measured by cross entropy. In the supervised training phase, the cross entropy loss function is defined as: ; in, is the true result distribution, Output probability distribution for prediction.

[0031] Back propagation and parameter update: Back propagate the loss function and calculate the gradient of each layer parameter. Based on the obtained gradient information, use the Adam optimizer to iteratively update the model parameters, with the goal of gradually reducing the value of the overall loss function, so that the model output is closer to the actual user preference.

[0032] In the SEAR method, supervised learning plays the following roles: Ensure initial recommendation capabilities: In the first stage of self-game, supervised learning is used to train a preliminary model based on real user data so that the model can generate reasonable recommendation results. Supervised learning ensures that the model has a relatively accurate recommendation strategy before starting self-game by minimizing the loss function.

[0033] Maintaining consistency in recommendation distribution: Supervised learning ensures that the output of the model matches the distribution of the training data, which is the prerequisite for direct preference optimization. By allowing the model to fine-tune its weights through real data during the training phase, the consistency between the recommendations generated by the model and the actual needs and preferences of users is guaranteed. Without this layer of supervised learning, the model may deviate from the reasonable recommendation direction during the self-game process, affecting the recommendation effect.

[0034] Improve the stability of self-game: supervised learning provides a stable initial state for the model, avoiding the strategy instability that may occur at the beginning of the self-game. Through fine-tuning of supervised learning, the output of the model can avoid over-reliance on adversarial strategies and can be optimized more stably during the self-game stage.

[0035] S300, direct preference optimization link, based on the preference data set generated by self-game, directly adjusts the recommendation strategy by optimizing the objective function, suppresses the over-recommendation of popular items and balances the long-tail distribution, wherein the optimization objective function is constructed by comparing the output probability difference between the current model and the reference model on positive and negative samples. It should be noted that the direct preference optimization link further includes generating diversified preference data through the self-game mechanism to increase the recommendation probability of low-frequency items.

[0036] It should be noted that direct preference optimization is one of the core links of the SEAR method, which aims to help the recommendation system better adapt to the real needs of users by directly optimizing the preferences of recommendation results. The core idea of ​​direct preference optimization is to use user preference data (i.e. which items are liked or ignored by users) to generate optimization targets, rather than relying on traditional label data. The specific steps are as follows: Acquisition of preference data sets: First, construct positive samples of preference data sets based on various user behaviors on the platform. Traditional methods usually adopt random sampling, semantic feature extraction and other principles to construct negative samples of preference data sets, while the SEAR method innovatively uses the self-game method to collect negative samples, solving the problem of bias that is difficult to solve in large model recommendation systems.

[0037] Forward propagation and loss calculation: Using the processed preference index, the optimization objective function that directly reflects the user's real interest is constructed as follows, that is, the optimization objective function of the direct preference optimization link is: ; in, For the current model, is the reference model, is the weight parameter; is a positive sample, is a negative sample, For input context.

[0038] This formula compares the current model With reference model In the positive sample and negative samples The output probability difference on is used to construct an optimization target that directly reflects the user’s true preference.

[0039] Back propagation and parameter update: Back propagate the loss function and calculate the gradient of each layer parameter. Based on the obtained gradient information, use the Adam optimizer to iteratively update the model parameters, with the goal of directly optimizing the real preferences of the target users.

[0040] The advantage of direct preference optimization is that it can optimize the recommendation strategy to enhance the fairness and diversity of the recommendation system. When combined with the self-game mechanism, direct preference optimization can further suppress the over-recommendation of popular items and balance the problem of long-tail distribution in the recommendation system. The self-game mechanism generates preference data by letting the model play against itself, which drives the model to continuously optimize the recommendation strategy in self-feedback.

[0041] The specific combination method is as follows: Self-game mechanism to generate data: In the data generation phase, the model generates recommendation data by playing against itself through a self-game mechanism. In each iteration, the model generates negative examples of recommended items based on the current strategy, and generates positive examples based on user preferences (such as clicks and ratings). These generated data include not only items preferred by users, but also negative items recommended by the model. The training data generated by self-game can better reflect the diversity of real-world recommendations and the limitations of model recommendation results.

[0042] Suppress popular items and balance long-tail distribution: Due to the high frequency of popular items in the recommendation system, traditional recommendation algorithms tend to favor popular items, resulting in the recommendation results being concentrated on a few items, resulting in the problem of "over-recommendation of popular items". The combination of the self-game mechanism and direct preference optimization can effectively avoid this problem. By generating diversified preference data through self-game, the model can balance the long-tail distribution, increase the recommendation probability of low-frequency items, and suppress over-recommendation of popular items.

[0043] The SEAR (debiasing recommendation method based on large language model based on self-game) proposed in the present invention has significant advantages and positive effects. First, SEAR generates training data in the training process through the self-game mechanism, avoiding the problem of excessive concentration on popular items in the current large language model recommendation system, thereby improving the diversity and fairness of the recommendation results. The adaptability of self-game enables the model to optimize the recommendation strategy through feedback without the intervention of external data, reducing dependence on external knowledge. Secondly, SEAR optimizes the ranking of recommended items in combination with direct preference optimization, further improving the recommendation accuracy and user satisfaction, balancing the recommendation of long-tail items and popular items, and enhancing the fairness and diversity of the system.

[0044] Empirical evaluations show that the deep integration of SEAR's self-game mechanism and the recommendation system makes it more flexible and adaptable when optimizing recommendation strategies, significantly improving the diversity, accuracy, and fairness of recommendations on real data sets.

[0045] In summary, SEAR provides an innovative solution to the debiasing problem of large language models in recommender systems, which has broad application potential and significant practical value.

[0046] In order to verify the effectiveness of the SEAR method, experiments were conducted on the MovieLens movie dataset, and BIGRec and the reweighting method were used as control models for comparison. The experiment used Llama2-7B as the base model and evaluated the model performance on two indicators: recommendation diversity and recommendation fairness. At the same time, the impact of introducing or not introducing statistical information in the recommendation stage on the recommendation results was investigated. Through this setting, the experiment aims to show the advantages of the SEAR method in improving recommendation diversity and fairness.

[0047] Table 1 Experimental results (top 5 recommended results)

[0048] As can be seen from Table 1, SEAR outperforms BIGRec and reweighted methods in terms of recommendation diversity and recommendation fairness. Without introducing statistical information, SEAR significantly improves recommendation diversity, while showing a significant reduction in average inter-group unfairness and maximum inter-group unfairness. Even with the introduction of statistical information, SEAR still shows significant improvements, especially in recommendation diversity. Overall, SEAR outperforms the control model in all indicators, proving that it effectively enhances the diversity and fairness of recommendations through self-game mechanisms and debiasing strategies.

[0049] from Figure 4 It can be seen that during the training process, the diversity of recommendations gradually increases, indicating that as the training progresses, the model is gradually able to generate more diverse recommendation results. This trend reflects that the model has gradually gotten rid of the over-recommendation of popular items in the process of continuous optimization, thereby improving the diversity of recommendations. In terms of unfairness, the experimental results show a clear downward trend, especially in the initial training rounds, when the model significantly reduced the unfairness of recommendations. Although the decline in unfairness becomes less obvious in subsequent training rounds, the overall trend still shows that the model is able to gradually balance the fairness of recommendations between different user groups. In general, the SEAR model shows a stable improvement in diversity and fairness during the training process, proving the effectiveness of the self-game mechanism in optimizing the diversity and fairness of the recommendation system.

[0050] The self-game-based large language model debiasing recommendation method (SEAR) proposed in the present invention can be widely used in various recommendation systems, especially in personalized recommendation tasks in the fields of movies, music, e-commerce, etc. For example, in the movie recommendation system, SEAR can effectively avoid over-recommendation of popular movies through the self-game mechanism, and improve the diversity and fairness of the recommendation results. Without relying on external data or complex post-processing, SEAR optimizes the recommendation strategy through a self-feedback mechanism, so that the recommendation system can provide users with more personalized and fair recommendation content, which is suitable for a variety of recommendation scenarios with long-tail effects and preference problems.

[0051] The embodiment of the present invention also provides a fair recommendation system based on large model self-game, which is used to implement the above method embodiment, which has been described and will not be repeated. The terms "module", "unit", "subunit", etc. used below can implement a combination of software and / or hardware for predetermined functions. Although the system described in the following embodiments is preferably implemented in software, the implementation of hardware or a combination of software and hardware is also possible and conceivable.

[0052] like Figure 5 As shown, Figure 5 It is a structural block diagram of a fair recommendation system based on large model self-game in the present invention, which includes: The self-game data generation module 101 is configured to: construct a preference data set based on user historical behavior data as positive samples and recommendation results generated by the current model recommendation strategy as negative samples, and remove samples with too high similarity between positive and negative samples through data filtering; The supervised training module 102 is configured to: perform supervised training using the preference data set based on the pre-trained large language model, optimize model parameters through cross entropy loss, and generate a preliminary recommendation strategy; The direct preference optimization module 103 is configured to: based on the preference data set generated by self-game, directly adjust the recommendation strategy by optimizing the objective function, suppress the over-recommendation of popular items and balance the long-tail distribution, wherein the optimization objective function is constructed by comparing the output probability difference between the current model and the reference model on positive and negative samples.

[0053] Figure 6 A structural block diagram of an electronic device provided by an embodiment of the present invention, such as Figure 6As shown, the electronic device may include: a processor 610, a communication interface 620, a memory 630 and a communication bus 640, wherein the processor 610, the communication interface 620 and the memory 630 communicate with each other through the communication bus 640. The processor 610 may call the logic instructions in the memory 630 to execute the following method: In the self-game data generation phase, based on the user's historical behavior data as positive samples and the recommendation results generated by the current model recommendation strategy as negative samples, a preference data set is constructed, and samples with too high similarity between positive and negative samples are removed through data filtering; In the supervised training phase, based on the pre-trained large language model, the preference data set is used for supervised training, the model parameters are optimized through cross entropy loss, and a preliminary recommendation strategy is generated; In the direct preference optimization phase, based on the preference data set generated by self-game, the recommendation strategy is directly adjusted by optimizing the objective function to suppress over-recommendation of popular items and balance the long-tail distribution. The optimization objective function is constructed by comparing the output probability difference between the current model and the reference model on positive and negative samples.

[0054] In addition, the logic instructions in the above-mentioned memory 630 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0055] An embodiment of the present invention further provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the method provided in the above embodiments is implemented.

[0056] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each embodiment or some parts of the embodiment.

[0057] 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 recommendation method based on large model self-game, characterized in that: The fair recommendation method based on large model self-game includes: Self-game data generation: Based on the user's historical behavior data as positive samples, the recommendation results generated by the current model recommendation strategy are used as negative samples to build a preference data set, and samples with too high similarity between positive and negative samples are removed through data filtering; Supervised training: Based on the pre-trained large language model, supervised training is performed using the preference dataset, and model parameters are optimized through cross entropy loss to generate a preliminary recommendation strategy. Direct preference optimization: Based on the preference data set generated by self-game, the recommendation strategy is directly adjusted by optimizing the objective function to suppress over-recommendation of popular items and balance the long-tail distribution. The optimization objective function is constructed by comparing the output probability difference between the current model and the reference model on positive and negative samples.

2. The fair recommendation method based on large model self-game according to claim 1, characterized in that: In the self-game data generation link, positive samples include user click, browse or rating behavior data; negative samples are generated through model iteration recommendation strategy and dynamically matched with user behavior data.

3. The fair recommendation method based on large model self-game according to claim 1, characterized in that: The data filtering includes: Calculate the Euclidean distance between the embedding vectors of the target recommended item and the predicted recommended item. If the distance is less than the preset threshold, the corresponding positive and negative samples are removed.

4. The fair recommendation method based on large model self-game according to claim 1, characterized in that: The model initialization of the supervised training phase uses the pre-trained large language model parameters, and the parameters are updated through the Adam optimizer.

5. The fair recommendation method based on large model self-game according to claim 1, characterized in that: In the supervised training phase, the cross entropy loss function is defined as: ； in, is the true result distribution, Output probability distribution for prediction.

6. The fair recommendation method based on large model self-game according to claim 1, characterized in that: The optimization objective function of the direct preference optimization link is: ； in, For the current model, is the reference model, is the weight parameter; is a positive sample, is a negative sample, For input context.

7. The fair recommendation method based on large model self-game according to claim 1, characterized in that: The direct preference optimization step further includes generating diversified preference data through a self-game mechanism to increase the probability of recommending low-frequency items.

8. A fair recommendation system based on large model self-game, characterized in that: include: The self-game data generation module is configured to: build a preference data set based on user historical behavior data as positive samples and the recommendation results generated by the current model recommendation strategy as negative samples, and remove samples with too high similarity between positive and negative samples through data filtering; The supervised training module is configured to: perform supervised training using the preference data set based on the pre-trained large language model, optimize model parameters through cross entropy loss, and generate a preliminary recommendation strategy; The direct preference optimization module is configured to: directly adjust the recommendation strategy by optimizing the objective function based on the preference data set generated by self-game, suppress the over-recommendation of popular items and balance the long-tail distribution, wherein the optimization objective function is constructed by comparing the output probability difference between the current model and the reference model on positive and negative samples.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, a fair recommendation method based on large model self-game as described in any one of claims 1 to 7 is implemented.

10. A non-transitory 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 fair recommendation method based on large model self-game as described in any one of claims 1 to 7 are implemented.

Images