Unbiased recommendation method based on implicit feedback attribution
By constructing a preference module, an exposure module, and an inference module, and using the expectation-maximization algorithm for alternating updates, the exposure bias problem in implicit feedback is solved, unbiased recommendation is achieved, and the performance of the recommendation system is improved.
Patent Information
- Application Number
- CN202211324146.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-10-27
- Publication Date
- 2026-01-23
- Estimated Expiration
- 2042-10-27
AI Technical Summary
Existing recommendation systems suffer from exposure bias when processing implicit feedback data, resulting in large model variance or biased estimations. They fail to fully utilize unclicked data and existing methods struggle to accurately distinguish between positive and negative signals.
An unbiased recommendation method based on implicit feedback attribution is adopted. By analyzing the four generation scenarios behind implicit feedback, a preference module, an exposure module, and an inference module are constructed. The expectation-maximization algorithm is used to update them alternately to extract reliable positive and negative signals.
It improves the performance of the recommendation system, enabling it to make full use of unclicked data and achieve efficient learning and recommendation of user interests and preferences.
Smart Images
Figure CN115618110B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of recommendation system rectification, in particular to the technical field of recommendation system rectification based on implicit feedback. BACKGROUND
[0002] Recommendation systems can provide personalized suggestions for users and play an important role in many online services. Current recommendation systems are usually built based on implicit feedback data (such as user clicks, etc.). Implicit feedback is a natural record of user behavior data, which is more abundant than explicit feedback (such as user ratings). However, due to the inherent exposure bias of implicit feedback, it is more challenging to process implicit feedback data. Exposure bias refers to the fact that implicit feedback is the joint result of user preference and exposure, and cannot fully represent user preference. Without addressing exposure bias, blindly fitting implicit feedback data will result in poor recommendation results.
[0003] To address the problem of exposure bias, existing methods mainly use methods such as de-weighting non-click data, modeling user exposure, and inverse propensity weighting. However, these methods have the problem of large model variance or biased model estimation, and do not fully utilize non-click data.
[0004] In implicit feedback, non-click data comes from the following three situations: 1) the user likes but does not know an item (positive signal); 2) the user knows but does not like an item (negative signal); 3) the user neither likes nor knows an item (negative signal). Non-click data contains both positive and negative signals, but existing methods either only utilize part of the signal or incorrectly treat positive signals as negative signals. To overcome this problem and extract reliable positive and negative signals, the reasons behind implicit feedback must be explicitly inferred. SUMMARY
[0005] To overcome the above problems, the present application proposes a new unbiased recommendation method based on implicit feedback attribution.
[0006] The present application can identify the three situations behind non-click data - unexposure, dislike, or both, thereby extracting reliable positive and negative signals from non-click data to learn accurate user interest preferences and achieve efficient recommendations.
[0007] The unbiased recommendation method based on implicit feedback attribution of the present application includes the following steps:
[0008] 1) Analyze and determine the four situations behind implicit feedback.
[0009] 2) Model user preference and exposure based on implicit feedback, and build an inference model to infer the reasons behind user implicit feedback.
[0010] 3) Alternating updates of the preference / exposure module and the inference module using the expectation-maximization algorithm.
[0011] 4) Using the trained user preference model to make predictions for the user's favorite items, and complete the recommendation.
[0012] Preferably, the four generating situations behind the implicit feedback in step 1) are analyzed as follows:
[0013] Assume y ui represents whether the user u clicks on the item i, r ui represents whether the user u is interested in the item i, o ui represents whether the user u has been exposed to the item i.
[0014] When the user's click on an item is not observed, i.e. y ui = 0, it can be due to the following three situations:
[0015] Case 1: r ui = 1 and o ui = 0, the user likes but does not know the item;
[0016] Case 2: r ui = 0 and o ui = 1, the user knows but does not like the item;
[0017] Case 3: r ui = 0 and o ui = 0, the user neither knows nor likes the item.
[0018] When the user's click on an item is observed, i.e. y ui = 1, there are:
[0019] Case 4: r ui = 1 and o ui = 1, the user both knows and likes the item.
[0020] Preferably, the preference module, exposure module and inference module in step 2) are constructed as follows:
[0021] The preference module (PM) is used to capture user preferences, which can be implemented using existing methods (such as matrix factorization). Given the features of the user-item pair (x u , x i ), the preference module will output the prediction of the user's interest in the item: β ui = p(r ui = 1|x u , x i ; M r ), where M rparameters of the preference module.
[0022] Exposure Module (EM) is used to estimate the likelihood of a user contacting an item: θ ui = p(o ui | x u , x i ; M o ).
[0023] Reasoning Module (RM) infers the reason of implicit feedback (one of the four). Given the collected data (y ui ) and the prior predictions of the preference module and exposure module (β ui , θ ui ), the reasoning module infers the posterior distribution of the preference (r ui ) and exposure (o ui ), i.e. the distribution of each user-item pair belonging to the four cases of implicit feedback. The reasoning module outputs a vector Each dimension represents the probability of belonging to each case, respectively:
[0024] Further, the implementation of the exposure module is: based on the item-based exposure model, assuming that the exposure probability of a user only depends on the characteristics of the item itself. This model simply introduces a learnable scalar parameter μ i that depends on the item, i.e. the predicted exposure probability θ ui = σ(μ i ), where σ(.) is the sigmoid function.
[0025] Alternatively, based on the matrix factorization exposure model, considering that the exposure probability depends on both the user and the item, different users may have different preferences and experience different system recommendation strategies for the same item, so the exposure probability for the same item may be different. The scalar based on the item is not sufficient to model the exposure, so this method proposes to use the latent factors of users and items to model the exposure, and to express the exposure probability as the dot product of the latent factors of users and items, i.e. where e u , e i , b u , b i are learnable parameters related to users and items.
[0026] The specific process of updating the preference / exposure module and the reasoning module alternately in step 3) by using the expectation maximization algorithm is as follows:
[0027] (3.1) Fix the parameters of the preference module and the exposure module, and update the reasoning module based on the Bayes formula:
[0028] p(r ui ,o ui |y ui M r M o )∝p(y ui |r ui ,o ui )p(r ui |M r )p(o ui |M o If ), then Φ is updated as follows:
[0029]
[0030] (3.2) Based on the posterior distribution obtained by the inference module We update the parameters of the preference module and exposure module, and optimize the expected probabilities as follows based on the EM algorithm:
[0031]
[0032] After simplifying the equations, we obtain the loss functions corresponding to the preference module and the exposure module:
[0033]
[0034] in The distance function representing the predicted value and the label r∈{0,1} can be expressed using cross-entropy loss.
[0035] Step 4) uses the trained user preference model to predict the items the user likes. The specific process is as follows: This method uses the observed implicit feedback biased dataset as the training set, and unbiased data collected from items randomly exposed with the same probability as the validation and test sets. Every 25 iterations, the performance of the current preference model is confirmed on the validation set. If the performance does not improve after a certain number of iterations, training ends.
[0036] The working principle of this invention is as follows: a preference module and an exposure module are introduced to capture user preferences and exposure patterns, and an inference module is introduced to reason about the reasons behind implicit feedback; simultaneously, an expectation-maximization algorithm is designed to alternately optimize the preference / exposure module and the inference module. Based on this, on the one hand, the preference and exposure modules can obtain reliable positive and negative sample signals from the inference module; on the other hand, the inference module can, in turn, use the information from the preference / exposure module to achieve more accurate inference.
[0037] The advantages of this invention are: by inferring the reasons behind implicit feedback, the model can make full use of unclicked data, improving recommendation performance, while also possessing good theoretical properties. Attached Figure Description
[0038] Figure 1 is the overall flowchart of the present application.
[0039] Figure 2 is the framework diagram of model training of the present application. DETAILED DESCRIPTION
[0040] The present application is further illustrated with reference to the accompanying drawings:
[0041] The implicit feedback attribution-based unbiased recommendation method of the present application comprises the following steps:
[0042] 1) analyzing and determining four generating conditions behind the implicit feedback.
[0043] 2) modeling the user preference and exposure based on the implicit feedback respectively, and constructing an inference model to infer the reasons for the user implicit feedback.
[0044] 3) alternately updating the preference / exposure module and the inference module using the expectation maximization algorithm.
[0045] 4) using the trained user preference model to predict the user's favorite items, and completing the recommendation.
[0046] The four generating conditions behind the implicit feedback in the step 1) are as follows:
[0047] Suppose y ui represents whether the user u clicks the item i, r ui represents whether the user u is interested in the item i, o ui represents whether the user u has contacted the item i.
[0048] When the user's click on a certain item is not observed, i.e. y ui = 0, it may be due to the following three conditions:
[0049] Condition 1: r ui = 1 and o ui = 0, the user likes but does not know the item;
[0050] Condition 2: r ui = 0 and o ui = 1, the user knows but does not like the item;
[0051] Condition 3: r ui = 0 and o ui = 0, the user neither knows nor likes the item.
[0052] When the user's click on a certain item is observed, i.e. y ui = 1, there are:
[0053] Condition 4: rui = 1 and o ui = 1, the user knows and likes the item.
[0054] The preference module, exposure module and inference module in step 2) are constructed as follows:
[0055] The preference module (PM) is used to capture user preferences, which can be implemented by existing methods (e.g., matrix factorization). Given the features of user-item pairs (x u , x i ), the preference module outputs the prediction of the user's interest level for the item: β ui = p(r ui = 1 | x u , x i ; M r ), where M r represents the parameters of the preference module.
[0056] The exposure module (EM) is used to estimate the likelihood of a user being exposed to an item: θ ui = p(o ui = 1 | x u , x i ; M o ). This method will give the following two implementations of the exposure module:
[0057] The first way: item-based exposure model, which assumes that the user exposure probability only depends on the characteristics of the item itself. This model simply introduces a learnable scalar parameter μ i , i.e., the predicted exposure probability θ ui = σ(μ i ), where σ(.) is the sigmoid function.
[0058] The second way: matrix factorization-based exposure model, which considers that the exposure probability depends on both the user and the item. Different users may have different preferences for the same item and experience different system recommendation strategies, so the exposure probability for the same item may be different. The scalar based on the item is not enough to model the exposure, so this method proposes to use the user and item hidden factors to model the exposure, and the exposure probability is expressed as the dot product of the user and item hidden factors, i.e. where e u , e i , b u , b i are learnable parameters related to users and items.
[0059] Reasoning Module (RM), which infers the reason of implicit feedback (one of the four). Given the collected data (y ui ) and the prior prediction of the Preference Module, Exposure Module (β ui , θ ui ), the Reasoning Module infers the posterior distribution of the preference (r ui ), exposure (o ui ), i.e., the distribution of each user-item pair belonging to the four cases of implicit feedback. The Reasoning Module outputs a vector Each dimension represents the probability of belonging to each case, respectively:
[0060] The specific process of updating the Preference / Exposure Module and the Reasoning Module alternately in step 3) by the Expectation-Maximization algorithm is as follows:
[0061] (3.1) Fix the parameters of the Preference Module and the Exposure Module, and update the Reasoning Module based on the Bayes formula:
[0062] p(r ui ,o ui |y ui ,M r ,M o )∝p(y ui |r ui ,o ui )p(r ui |M r )p(o ui |M o ), then update Φ as follows:
[0063]
[0064] (3.2) Update the parameters of the Preference Module and the Exposure Module according to the posterior distribution of the Reasoning Module. We optimize the expected probability as follows according to the EM algorithm:
[0065]
[0066] After simplifying the equation, we get the loss functions corresponding to the Preference Module and the Exposure Module:
[0067]
[0068] where represents the distance function between the predicted value and the label r∈{0,1}, which can use cross-entropy loss.
[0069] The specific process of using the trained user preference model to make a prediction for the user's favorite item in step 4 is as follows: the method takes the observed implicit feedback biased data set as the training set, and the unbiased data collected on the same probability exposure items as the validation set and test set. Every 25 iterations, the performance of the current preference model is confirmed on the validation set, and when the performance does not improve after a certain number of iterations, the training is ended.
[0070] The present application proposes an unbiased recommendation method based on implicit feedback attribution. The method is composed of a preference module, an exposure module and an inference module. The preference module and the exposure module can extract reliable signals from the inference module, and the inference module can also use the information in the preference module and the exposure module. The method solves the exposure bias problem by inferring the cause of implicit feedback, and realizes unbiased recommendation. Compared with other existing methods, the method can fully utilize the non-click data, improve the performance of recommendation, and has good theoretical properties.
[0071] The content described in the embodiments of the present application is only a list of implementation forms of the inventive concept, and the protection scope of the present application should not be regarded as limited to the specific forms stated in the embodiments, and the protection scope of the present application also extends to equivalent technical means that can be thought of by those skilled in the art according to the inventive concept.
Claims
1. An unbiased recommendation method based on implicit feedback attribution, characterized in that: Includes the following steps: 1) Analyze and identify the four possible scenarios that generate implicit feedback; 2) Based on implicit feedback, model user preferences and exposure separately, and construct inference models to reason about the reasons for users' implicit feedback; specifically including: The Preference Module (PM) is used to capture user preferences, given the characteristics of user-item pairs. The preference module will output a prediction of the user's level of interest in the items: ,in Indicates the parameters of the preference module; The Exposure Module (EM) is used to estimate the likelihood of a user coming into contact with an object. ; The Reasoning Module (RM) infers the reasons for implicit feedback; given the collected data... Prior predictions for preference and exposure modules The inference module infers preferences. ,exposure The posterior distribution, i.e., the distribution of each user-item pair belonging to one of the four implicit feedback scenarios; the inference module will output a vector. Each dimension represents the probability of belonging to each case: ; 3) The preference / exposure module and the inference module are updated alternately using the expectation-maximization algorithm; the specific process is as follows: (1) The parameters of the fixed preference module and the exposure module are updated in the inference module based on Bayes' theorem: Then update as follows: (2) Based on the posterior distribution obtained from the inference module Update the parameters of the preference module and exposure module, and optimize the expected probabilities as follows based on the EM algorithm: After simplifying the equations, we obtain the loss functions corresponding to the preference module and the exposure module: in Indicates predicted value and label The distance function can be obtained using cross-entropy loss; 4) Use the trained user preference model to predict the items that users like and make recommendations.
2. The unbiased recommendation method based on implicit feedback attribution as described in claim 1, characterized in that: Step 1) analyzes the four possible scenarios behind implicit feedback as follows: Assumption Indicates user Did you click on the item? , Indicates user Is it for the item? interested, Indicates user Did you touch the item? ; When no user clicks on a particular item are observed, i.e. This could stem from one of the following three situations: Scenario 1: and Users like the item but are unaware of it; Scenario 2: and The user knows about the item but doesn't like it; Scenario 3: and The user neither knows nor likes the item; When a user clicks on an item, that is ,have: Scenario 4: and Users both know about and like the item.
3. The unbiased recommendation method based on implicit feedback attribution as described in claim 1, characterized in that: The exposure module is implemented as follows: based on an item-based exposure model, it assumes that the user's exposure probability depends only on the characteristics of the item itself, and introduces a learnable scalar parameter that depends on the item. That is, the predicted exposure probability ,in This is the sigmoid function.
4. The unbiased recommendation method based on implicit feedback attribution as described in claim 1, characterized in that: The exposure module is implemented as follows: Based on a matrix factorization-based exposure model, the exposure probability is considered to depend on both the user and the item. Different users may have different preferences for the same item and experience different system recommendation strategies, thus the exposure probability for the same item may vary. Since scalar values based on the item are insufficient for modeling exposure, latent factors of the user and item are used to model exposure, expressing the exposure probability as the dot product of the latent factors of the user and the item. ,in These are learnable parameters related to users and items.
5. The unbiased recommendation method based on implicit feedback attribution as described in claim 1, characterized in that... Step 4) uses the trained user preference model to predict the items the user likes. The specific process is as follows: the observed implicit feedback biased dataset is used as the training set, and the unbiased data collected from items randomly exposed with the same probability is used as the validation set and test set; every 25 iterations, the performance of the current preference model is confirmed on the validation set, and training ends when the performance does not improve after a certain number of iterations.