Unbiased recommendation method based on auxiliary information learning invariance and environmental preference

Abstract

The invention discloses an unbiased recommendation method based on auxiliary information learning invariance and environment preference in the technical field of causal invariance and unbiased recommendation, which comprises the following steps of: obtaining interaction data of a user and a project according to a causal relationship graph, and inputting auxiliary information in the data into an environment classifier; performing environment division on the input auxiliary information by taking the auxiliary information as guidance; respectively learning invariant preferences and environment specific preferences based on environment division so as to capture the invariant preferences and the environment specific preferences of the user; and outputting a final prediction result based on the learned invariant preferences of the user and the environment specific preferences of the environment to which the user belongs. According to the method, more reliable environment division is obtained through guidance of the auxiliary information, the inaccuracy of environment division when invariant preferences are learned in the past is overcome, specific environment preferences are added on the basis, dynamic changes of user preferences in the environment are considered, and the accuracy of unbiased recommendation and the dynamism of the environment preferences are improved.

Description

Technical Field

[0001] The present invention relates to the field of causal invariance and unbiased recommendation technology, and in particular to an unbiased recommendation method based on auxiliary information learning invariance and environmental preference. Background Art

[0002] How to learn the unbiased preferences of users is the fundamental problem of recommendation. The mixed data distribution in heterogeneous data will hinder the recommendation system from capturing the real preferences of users, thus leading to biased recommendations. The current debiasing recommendation technologies mainly include: ① debiasing method based on IPS; ② debiasing method based on information bottleneck; ③ debiasing method based on invariant learning.

[0003] Among the three debiasing methods mentioned above, the debiasing method based on invariant learning is the most widely used. The other methods can only remove certain specific biases, while the method based on invariant learning can remove general biases. It captures the user's unbiased preferences based on the latest causal invariance theory to achieve debiasing.

[0004] Although the current invariant learning recommendation method is committed to eliminating bias and improving the stability of recommendations, it still has multiple problems and defects: 1. Inaccurate environment division: The general debiasing framework of the existing methods does not have any guarantee for the environment division, and its environment division is inaccurate.

[0005] 2. Poor algorithm: The existing methods use a step-by-step iterative algorithm for environmental division, and the algorithm has a high time complexity.

[0006] 3. Ignoring specific environmental preferences: Existing methods only use invariant preferences across environments for prediction, without considering preferences in specific environments. They only focus on achieving multi-objective optimization in multiple environments, while damaging the accuracy in a single environment and the dynamics of environmental preferences. Summary of the invention

[0007] The purpose of this section is to summarize some aspects of the embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the specification abstract and the invention title of this application to avoid blurring the purpose of this section, the specification abstract and the invention title, and such simplifications or omissions cannot be used to limit the scope of the present invention.

[0008] Therefore, the purpose of the present invention is to provide an unbiased recommendation method based on auxiliary information learning invariance and environmental preference, which improves the accuracy of unbiased recommendation and the dynamics of environmental preference.

[0009] To solve the above technical problems, according to one aspect of the present invention, the present invention provides the following technical solutions: An unbiased recommendation method based on auxiliary information learning invariance and environmental preference, the specific steps are as follows: S1. Obtain the data of user and item interaction based on the causal relationship diagram; S2, input the auxiliary information in the data into the environment classifier, and use the auxiliary information as a guide to perform environment classification on the input auxiliary information; S3, learning invariant preferences and environment-specific preferences based on environment partitioning to capture users’ invariant preferences and environment-specific preferences; S4. Output the final prediction result based on the learned invariant preferences of the user and the environment-specific preferences of the environment to which the user belongs.

[0010] As a preferred solution of the unbiased recommendation method based on auxiliary information learning invariance and environmental preference described in the present invention, in step S2, the formula for environmental division of the input auxiliary information using the auxiliary information as a guide is as follows: ; ; ; in, and are two linear layers, A is auxiliary information, is the softmax function, express The kth item of , outputs h(A) as the probability distribution of the environment category, ensuring that the sum of the probabilities of all environment categories is 1.

[0011] As a preferred solution of the unbiased recommendation method based on auxiliary information learning invariance and environmental preference described in the present invention, in step S3, the steps of learning invariant preferences based on environmental division are as follows: With two embeddings and The Hadamard product models the invariant preference of user u for item i To make It can capture the user's invariant preferences across environments. The formula is as follows: ; in, and Represent the invariant embeddings of user u and item i respectively; Define the loss function The predicted value can be converted into With actual value For comparison, the formula is as follows: ; ; in, is the feedback function, is the cross entropy loss; Introducing a constant penalty , the announcement is as follows: .

[0012] As a preferred solution of the unbiased recommendation method based on auxiliary information learning invariance and environmental preference described in the present invention, in step S3, the step of learning environment-specific preferences based on environment division to capture the user's environment-specific preferences is as follows: use and To represent the context-specific embedding of user u and item i in the kth context, we use and The Hadamard product models the specific context preference of user u for item i , the formula is as follows: ; ; in, Indicate environmental preferences; Define the loss function To capture the final feedback of user u on item i, the formula is as follows: ; ; in, is the feedback function, is the cross entropy loss, is the actual value, represents the interaction between user u and item i in environment e; Minimize the following objective function: ; in and is a hyperparameter.

[0013] Compared with the prior art, the present invention has the following beneficial effects: the present invention obtains more reliable environment division by using auxiliary information as guidance, overcomes the inaccuracy of environment division when learning invariant preferences in the past, and adds environment-specific preferences on this basis, taking into account the dynamic changes of user preferences in the environment, and achieves better performance on the basis of previous debiasing models, thereby improving the accuracy of unbiased recommendations and the dynamics of environmental preferences. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below in combination with the accompanying drawings and detailed embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. Among them: Figure 1 The present invention is a flowchart of an unbiased recommendation method based on auxiliary information learning invariance and environmental preference.

[0015] Figure 2 A cause-and-effect diagram provided for the present invention.

[0016] Figure 3 The present invention is a flowchart of the practical application of an unbiased recommendation method based on auxiliary information learning invariance and environmental preference. DETAILED DESCRIPTION

[0017] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0018] Secondly, the present invention is described in detail with reference to schematic diagrams. When describing the embodiments of the present invention in detail, for the sake of convenience, the cross-sectional diagrams showing the device structure will not be partially enlarged according to the general scale, and the schematic diagrams are only examples, which should not limit the scope of protection of the present invention. In addition, in actual production, the three-dimensional dimensions of length, width and depth should be included.

[0019] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0020] The present invention provides an unbiased recommendation method based on auxiliary information learning invariance and environmental preference, which improves the accuracy of unbiased recommendation and the dynamics of environmental preference.

[0021] First, according to Figure 2 The causal relationship diagram generated by the data shown in the figure shows that the interaction between users and items is obtained by the user's invariant preferences and the user's environment-specific preferences. Then, the auxiliary information in the input data is input into the environment classifier, and the environment of the input data is divided by the auxiliary information as a guide. Then, through the joint participation of the environment classifier, the invariant preference and environment-specific preference modules are modeled respectively to capture the user's invariant preferences and environment-specific preferences. Finally, predictions are made based on the user's invariant preferences and the environment-specific preferences of the user's environment, such as Figure 1 As shown, the specific steps are as follows: Step 1: Environment Classifier Module Considering that there is additional auxiliary information A∈ in the set (U, I, Y) of interactions between U, item I and user-item , proposed an environment classifier that uses auxiliary information A for environment reasoning and performs invariant learning from a heterogeneous dataset D without environment partitioning. The formula is as follows: (1) ; ; in, and are two linear layers, is the softmax function, express The kth item of . Output h(A) as the probability distribution of the environment category, ensuring that the sum of the probabilities of all environment categories is 1 (i.e., the second column of formula (1)). This stage effectively identifies the user's environment and provides the necessary environment information for subsequent modeling.

[0022] Step 2: Learning invariant preferences With two embeddings and The Hadamard product models the invariant preference of user u for item i , the formula is as follows: (2) in, and Represent the invariant embeddings of user u and item i respectively. The goal of this step is to make Able to capture users' invariant preferences across environments.

[0023] In order to achieve the above goal, a loss function is defined Come let Capturing the invariant preference, through the feedback function The predicted value With actual value For comparison, the formula is as follows: (3) (4) in is the feedback function, is the cross entropy loss.

[0024] To ensure The captured preferences are invariant across environments, introducing a constant penalty term To ensure the consistency of the invariant preference in different environments, the environmental classifier is required in this step. The hope is that the invariant preference has the best performance in all environments, and an optimal preference specific to the environment is fitted on the data of the environment. ,if Preference for unchanged To reach a smaller value, then we know It is not optimal in this environment, so further training is done to optimize the environment-specific preferences. To measure invariant preference Whether it is optimal in all environments, the constant penalty term is given as follows: (5) The above formula is somewhat challenging. Replace the constant penalty term with its first-order approximation, and the formula is as follows: (6) By introducing the penalty term, we can ensure that the model will not overfit to a specific environment during training and maintain the stability of the constant preference.

[0025] Step 3: Learning environment-specific preferences use and to represent the context-specific embedding of user u and item i in the kth context, and similarly use their Hadamard product to model the context-specific preference of user u for item i , the formula is as follows: (7) (8) Formula (7) represents the environmental preference Capture each data's environment-specific preferences , which combines all environment-specific preferences into one set.

[0026] Next, we define a loss function To capture the final feedback of user u on item i, the formula is as follows: (9) (10) in is the feedback function, is the cross entropy loss.

[0027] The ultimate goal is to minimize the following objective function: (11) in and These are hyperparameters. By adjusting these hyperparameters, we can flexibly control the impact of each loss term on model training.

[0028] The above unbiased recommendation method based on auxiliary information learning invariance and environmental preference is used in practical applications, such as Figure 3 As shown in the figure, the details are as follows: based on the input data, the environment to which the input data belongs is first determined by the environment classifier, and the environment-specific preference module corresponding to the environment is found through the judgment result of the environment classifier, and the final prediction result is output in combination with the invariant preference module. The application flow chart is shown in Figure 3 shown.

[0029] Although the present invention has been described above with reference to the embodiments, various modifications may be made thereto and parts thereof may be replaced by equivalents without departing from the scope of the present invention. In particular, as long as there is no structural conflict, the various features in the embodiments disclosed in the present invention may be used in combination with each other in any manner, and the fact that these combinations are not exhaustively described in this specification is only for the sake of omitting space and saving resources. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. An unbiased recommendation method based on auxiliary information learning invariance and environmental preference, characterized in that: The specific steps are as follows: S1. Obtain the data of user and item interaction based on the causal relationship diagram; S2, input the auxiliary information in the data into the environment classifier, and use the auxiliary information as a guide to perform environment classification on the input auxiliary information; S3, learning invariant preferences and environment-specific preferences based on environment partitioning to capture users’ invariant preferences and environment-specific preferences; S4. Output the final prediction result based on the learned invariant preferences of the user and the environment-specific preferences of the environment to which the user belongs.

2. The unbiased recommendation method based on auxiliary information learning invariance and environmental preference according to claim 1, characterized in that: In step S2, the formula for dividing the environment of the input auxiliary information using the auxiliary information as a guide is as follows: ； ； ； in, and are two linear layers, A is auxiliary information, is the softmax function, express The kth item of , outputs h(A) as the probability distribution of the environment category, ensuring that the sum of the probabilities of all environment categories is 1.

3. The unbiased recommendation method based on auxiliary information learning invariance and environmental preference according to claim 1, characterized in that: In step S3, the steps of learning the invariant preferences based on the environment division are as follows: With two embeddings and The Hadamard product models the invariant preference of user u for item i To make It can capture the user's invariant preferences across environments. The formula is as follows: ； in, and Represent the invariant embeddings of user u and item i respectively; Defining the loss function The predicted value can be converted into With actual value For comparison, the formula is as follows: ； ； in, is the feedback function, is the cross entropy loss; Introducing a constant penalty , the announcement is as follows: 。 4. The unbiased recommendation method based on auxiliary information learning invariance and environmental preference according to claim 1, characterized in that: In step S3, the steps of learning environment-specific preferences based on environment segmentation to capture the user's environment-specific preferences are as follows: use and To represent the context-specific embedding of user u and item i in the kth context, we use and The Hadamard product models the specific context preference of user u for item i , the formula is as follows: ； ； in, Indicate environmental preferences; Defining the loss function To capture the final feedback of user u on item i, the formula is as follows: ； ； in, is the feedback function, is the cross entropy loss, is the actual value, represents the interaction between user u and item i in environment e; Minimize the following objective function: ； in and is a hyperparameter.

Images