Fair recommendation method based on prejudice quantification and utility tradeoff

Abstract

The invention discloses a fair recommendation method based on prejudice quantification and utility tradeoff, which comprises the following steps of: automatically classifying or manually classifying items, and estimating a prejudice value of a user group to a certain type of items through quantification of scores of users to the items according to the user group and the item types; performing depolarization processing on the user score by using the prejudice value, generating a prejudice-free score, inputting the prejudice-free score into a recommendation algorithm, generating a prejudice-free prediction score, and taking the first N items to generate a recommendation list; constructing a user attention model, selecting a fairness index, and constructing a measurement strategy of effectiveness and fairness; exchanging recommended item positions in the ranking list, and calculating the measurement change of ranking effectiveness and fairness before and after exchange; and iteratively updating the recommendation list through an exchange mechanism, and generating an optimal balanced top-k recommendation list. According to the method, fairness recommendation is realized through data depolarization and utility tradeoff, and the problem of data prejudice and the problem of tradeoff of fairness and utility of a current recommendation system are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of recommendation systems, and in particular to a fair recommendation method based on bias quantification and utility weighing. Background Art

[0002] In the process of optimizing user experience and commercial benefits, existing recommendation systems often face fairness issues, that is, the recommendation results may be biased against certain user groups. These biases usually come from historical biases in the training data or limitations of the model itself, resulting in certain user groups being systematically ignored or treated unfairly. Traditional fair recommendation methods usually only focus on eliminating bias, but this may lead to a decrease in the effectiveness of recommendations. Therefore, how to maintain the effectiveness of the recommendation system while eliminating bias has become a key issue that needs to be solved urgently. Summary of the invention

[0003] In view of the above-mentioned shortcomings, the present invention proposes a fair recommendation method based on bias quantification and utility trade-off, which can achieve fair recommendation by debiasing data and weighing utility, and solve the data bias problem and the trade-off problem between fairness and utility of the current recommendation system.

[0004] To achieve the above object, the present invention provides the following technical solution: a fair recommendation method based on bias quantification and utility trade-off, comprising the following steps:

[0005] S1: Automatically classify items through online API services or manually classify items through crowdsourcing services, and estimate the bias value θ of the user group for a certain type of item based on the geometric mean of the user u's rating of item I according to the user group U and item category C u ;

[0006] S2: Use the bias value to debias the user rating and generate an unbiased rating d ui , input unbiased ratings into the recommendation algorithm to generate unbiased predicted ratings And take the first N items to generate a recommendation list;

[0007] S3: Build a user attention model, select appropriate fairness indicators, and build a measurement strategy for utility F and fairness G;

[0008] S4: Exchange the positions of the recommended items in the ranking list, and calculate the metric changes of utility F and fairness G according to the rankings before and after the exchange;

[0009] S5: Iteratively update the recommendation list through the exchange mechanism to generate the top-k recommendation list with the best trade-off.

[0010] As an improvement, the user groups in step S1 include a disadvantaged group D and an advantaged group A.

[0011] As an improvement, in step S1, for the disadvantaged group user U d For category c k Project Group The formula for the geometric mean of the ratings is as follows:

[0012]

[0013] For advantage group users U a For category c k Project Group The formula for the geometric mean of the ratings is as follows:

[0014]

[0015] Among them, r ui represents the rating of item i by user u.

[0016] As an improvement, the bias value θ of the user group for a certain type of item in step S1 u The calculation formula is as follows:

[0017]

[0018] As an improvement, in step S2, for user u with category c k The formula for calculating the debiased score of item i is as follows:

[0019]

[0020] Among them, r ui represents the rating of item i by user u.

[0021] As an improvement, in step S3, a reciprocal decay function is used to estimate the user's attention to construct a user attention model. The calculation formula is as follows:

[0022]

[0023] Among them, a m represents the attention value at the mth position, α is a parameter that controls the decay speed, larger α values ​​mean faster decay, and k is the number of the first k items.

[0024] As an improvement, the utility F in step S3 is calculated by accumulating the relevance score, and the calculation formula is as follows:

[0025]

[0026] In the above formula, a i is the user's attention to item i, is the predicted rating of user u for item i.

[0027] As an improvement, in step S3, the Gini index fairness metric is used for calculation, and the calculation formula is as follows:

[0028]

[0029] In the above formula, E i It is the cumulative exposure of the first i items in the recommendation list.

[0030] As an improvement, the utility change ΔF after exchanging the item positions in step S4 is calculated as follows:

[0031] ΔF=F new -F old

[0032] In the above formula, F new is the utility value after exchanging the position of the items, F old is the utility value before swapping the item positions,

[0033] The fairness change ΔG after exchanging the item position is calculated as follows:

[0034] ΔG=G new -G old

[0035] In the above formula, G new is the fairness value after exchanging the item positions, G old is the fairness value before swapping the item positions.

[0036] As an improvement, the exchange mechanism in step S5 is to determine whether to accept the current exchange based on a preset threshold. If ΔF and ΔG meet Then accept the exchange and update the current ranking, recalculate the changes in utility F and fairness G for the items after the recommended item position until reaching

[0037] Compared with the prior art, the advantages of the present invention are:

[0038] (1) The present invention quantifies the bias value of user groups towards specific item categories and performs debiasing processing during the scoring process. This method significantly reduces the unfairness of recommendations caused by user bias, which helps ensure that the recommendation system more objectively and fairly reflects the real interests and needs of users.

[0039] (2) By generating unbiased ratings and inputting them into the recommendation algorithm, this method can generate more accurate and reliable recommendation lists, which not only improves user satisfaction with the recommended content, but also enhances the overall performance and credibility of the recommendation system;

[0040] (3) When constructing the recommendation list, this method not only considers the recommendation utility (i.e., the user's satisfaction with the recommended items), but also takes fairness into account by constructing a user attention model and selecting appropriate fairness indicators. This dual consideration enables the recommendation results to meet the personalized needs of users while ensuring fairness and diversity among different user groups.

[0041] (4) By iteratively updating the recommendation list and accepting or rejecting the exchange of item positions according to the preset trade-off threshold, this method provides a high degree of flexibility and adjustability, which allows the system to adjust the balance between fairness and utility according to actual needs to adapt to different application scenarios and user preferences;

[0042] (5) By continuously optimizing the order and combination of items in the recommendation list, this method can improve the user experience. Users can not only obtain recommendations that are more in line with their interests and needs, but also feel that the recommendation system attaches importance to fairness.

[0043] (6) By paying special attention to the needs of disadvantaged groups, this method helps to increase the diversity and inclusiveness of recommended content. By reducing bias and discrimination, this method provides more equal and fair recommendation opportunities for different user groups. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0045] Figure 1 It is the overall flow chart of the method of the present invention;

[0046] Figure 2 is a flow chart of the bias quantification and debiasing recommendation module of the present invention;

[0047] Figure 3 It is a flow chart of the utility and fairness trade-off module of the present invention. DETAILED DESCRIPTION

[0048] like Figures 1 to 3As shown, a fair recommendation method based on bias quantification and utility trade-off includes a bias quantification module, a debiasing recommendation module, and a utility and fairness trade-off module, wherein the bias quantification module is used to estimate the bias value of the user group, specifically by automatically classifying the items through an online API service or manually classifying them through a crowdsourcing service, and estimating the bias value of the user group for a certain type of project through the geometric mean of the user's ratings of the project according to the user group and the project category; the debiasing recommendation module is used to generate unbiased prediction scores, specifically by debiasing the user ratings using the bias values ​​to generate unbiased scores, and the unbiased scores are input into the recommendation algorithm to generate unbiased prediction scores; the utility and fairness trade-off module is used to build a user attention model, specifically by selecting appropriate fairness indicators, building a utility and fairness measurement strategy, exchanging the positions of the recommended items in the ranking list, calculating the utility and fairness measurement changes according to the rankings before and after the exchange, and iteratively updating the recommendation list through the exchange mechanism to obtain the optimally weighed recommendation list.

[0049] The present invention is implemented under the condition that the recommendation results in the recommendation scenario pursue a balance between utility and fairness. In this embodiment, a recommendation model is constructed using a collaborative filtering algorithm, and recommendations are made based on user-movie rating data of a movie push platform.

[0050] In the following steps, U represents the user set, I represents the movie set, C represents the movie theme category, and the movie theme types include action, suspense, costume, romance, and comedy, etc., A represents the user set related to the dominant group, and D represents the user set related to the disadvantaged group. The dominant and disadvantaged groups are divided according to the number of male and female users who rate movies of the theme type (for example, among the users who rate love movies, there are more female users than male users, so female users belong to the dominant group and male users belong to the disadvantaged group). ui represents the rating of user u on movie i, θ u Represents the bias value of user u.

[0051] A fair recommendation method based on bias quantification and utility trade-off includes the following steps:

[0052] S1.1: Preprocess the data and classify the movies by genre through API services or crowdsourcing services.

[0053] S1.2: Quantify the bias and calculate the bias for disadvantaged group users U d For the subject category c k User groups The geometric mean of the ratings is:

[0054]

[0055] For advantage group users Ua For the subject category c k User groups The geometric mean of the ratings is:

[0056]

[0057] The bias value θ of user u u Calculation formula:

[0058]

[0059] S2.1: Use the bias value to debias the user's ratings of the movie and generate unbiased ratings.

[0060] For user u, the subject category is c k Debiased ratings of movie i:

[0061]

[0062] Among them, r ui represents the rating of user u on movie i, θ u Represents the bias value of user u.

[0063] S2.2: Recommendation generation, the debiased score d ui Input collaborative filtering algorithm to generate user u's predicted rating for movie i Take the first N items to generate a candidate recommendation list;

[0064] S3: Measuring utility and fairness. The utility F is calculated by accumulating the prediction score. The calculation formula is:

[0065]

[0066] Among them, a i is the user's attention to item i, is the predicted score.

[0067] Fairness G is calculated by fairness metrics such as the Gini index, and the calculation formula is:

[0068]

[0069] Among them, E i It is the cumulative number of views, collections, and plays of the first i movies in the recommendation list.

[0070] S4: Starting from the initial ranking, the candidate recommendation list is adjusted by swapping the positions of items i and j. Each swap affects the utility and fairness metrics.

[0071] After exchanging i and j, the utility change ΔF is calculated as:

[0072] ΔF=F new -F old

[0073] In the above formula, F new is the utility value after exchanging the position of the items, F old is the utility value before exchanging the positions of the items.

[0074] After exchanging i and j, the fairness change ΔG is calculated as:

[0075] ΔG=G new -G old

[0076] In the above formula, G new is the fairness value after exchanging the item positions, G old is the fairness value before swapping the item positions.

[0077] S5: Determine whether to accept the current exchange based on the preset threshold. If ΔF and ΔG meet If the exchange is accepted, the current ranking is updated and step S6 is repeated until the exchange is reached. Finally, a top-k recommendation list with optimal trade-offs is generated.

[0078] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present disclosure. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some implementations as replacements, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0079] The units described in some embodiments of the present disclosure may be implemented in software or hardware. The units described may also be arranged in a processor, and the functions described above may be at least partially performed by one or more hardware logic components.

[0080] The present invention is described above by way of example in conjunction with the accompanying drawings. It is obvious that the specific implementation of the present invention is not limited to the above-mentioned method. As long as various non-substantial improvements are made using the technical solution of the present invention, or the concept and technical solution of the present invention are directly applied to other occasions without improvement, they are all within the protection scope of the present invention.

Claims

1. A fair recommendation method based on bias quantification and utility trade-off, characterized in that: The following steps are involved: S1: Automatically classify items through online API services or manually classify items through crowdsourcing services, and estimate the bias value θ of the user group for a certain type of item based on the geometric mean of the user u's rating of item I according to the user group U and item category C u ; S2: Use the bias value to debias the user rating and generate an unbiased rating d ui , input unbiased ratings into the recommendation algorithm to generate unbiased predicted ratings And take the first N items to generate a recommendation list; S3: Build a user attention model, select appropriate fairness indicators, and build a measurement strategy for utility F and fairness G; S4: Exchange the positions of the recommended items in the ranking list, and calculate the metric changes of utility F and fairness G according to the rankings before and after the exchange; S5: Iteratively update the recommendation list through the exchange mechanism to generate the top-k recommendation list with the best trade-off.

2. The fair recommendation method based on bias quantification and utility trade-off according to claim 1, characterized in that: In step S1, the user groups include a disadvantaged group D and an advantaged group A.

3. The fair recommendation method based on bias quantification and utility trade-off according to claim 2, characterized in that: In step S1, for the disadvantaged group user U d For category c k Project Group The formula for the geometric mean of the ratings is as follows: For advantage group users U a For category c k Project Group The formula for the geometric mean of the ratings is as follows: Among them, r ui represents the rating of item i by user u.

4. The fair recommendation method based on bias quantification and utility trade-off according to claim 3, characterized in that: The bias value θ of the user group towards a certain type of item in step S1 u The calculation formula is as follows:

5. The fair recommendation method based on bias quantification and utility trade-off according to claim 1, characterized in that: In step S2, for user u, category c k The formula for calculating the debiased score of item i is as follows: Among them, r ui represents the rating of item i by user u.

6. The fair recommendation method based on bias quantification and utility trade-off according to claim 1, characterized in that: In step S3, a reciprocal decay function is used to estimate the user's attention to construct a user attention model, and the calculation formula is as follows: Among them, a m represents the attention value at the mth position, α is a parameter that controls the decay speed, larger α values ​​mean faster decay, and k is the number of the first k items.

7. The fair recommendation method based on bias quantification and utility trade-off according to claim 1, characterized in that: In step S3, the utility F is calculated by accumulating the correlation score, and the calculation formula is as follows: In the above formula, a i is the user's attention to item i, is the predicted rating of user u for item i.

8. The fair recommendation method based on bias quantification and utility trade-off according to claim 1, characterized in that: In step S3, the Gini index fairness metric is used for calculation, and the calculation formula is as follows: In the above formula, E i It is the cumulative exposure of the first i items in the recommendation list.

9. The fair recommendation method based on bias quantification and utility trade-off according to claim 1, characterized in that: The utility change ΔF after exchanging the item positions in step S4 is calculated by the following formula: ΔF=F new -F old In the above formula, F new is the utility value after exchanging the position of the items, F old is the utility value before swapping the item positions, The fairness change ΔG after exchanging the item position is calculated as follows: ΔG=G new -G old In the above formula, G new is the fairness value after exchanging the item positions, G old is the fairness value before swapping the item positions.

10. The fair recommendation method based on bias quantification and utility trade-off according to claim 9, characterized in that: The exchange mechanism in step S5 is to determine whether to accept the current exchange based on a preset threshold. If ΔF and ΔG meet Then accept the exchange and update the current ranking, recalculate the changes in utility F and fairness G for the items after the recommended item position until reaching

Images