Search

CN-121981274-A - Anti-facts recommendation method, device, equipment and storage medium based on causal inference

CN121981274ACN 121981274 ACN121981274 ACN 121981274ACN-121981274-A

Abstract

The invention relates to the technical field of artificial intelligence, which can be applied to the medical field and the financial science and technology field, and discloses a causal inference-based anti-facts recommendation method, device, equipment and storage medium, wherein the method comprises the steps of obtaining an observation data set generated by a recommendation system and constructing a causal graph; the method comprises the steps of taking user characteristics and article characteristics in an observation data set as input, estimating trend scores through a classification model, estimating average causal effects of exposure behaviors on user feedback by adopting a double machine learning mechanism, calculating potential anti-facts results, constructing a training data set based on the potential anti-facts results and the observation data set, training a depth recommendation model based on the training data set, predicting candidate articles through the depth recommendation model obtained through training to obtain target predicted values, and generating a recommendation list. The invention can effectively correct exposure deviation, selection deviation and confounding factor interference in the observed data, thereby providing a fairer, accurate and interpretable recommended result.

Inventors

  • DENG YUWEI
  • KONG LINGWEI

Assignees

  • 平安科技（深圳）有限公司

Dates

Publication Date
20260505
Application Date
20260302

Claims (10)

  1. 1. A causal inference-based anti-facts recommendation method, comprising: Acquiring an observation data set generated by a recommendation system, and constructing a causal graph based on potential relations between service logic of the recommendation system and variables in the observation data set; Based on the causal relationship in the causal graph, taking the user features and the object features in the observation data set as input, and estimating a tendency score through a classification model; estimating an average causal effect of exposure behavior on user feedback based on the observation data set and the trend score by adopting a double machine learning mechanism, and calculating a potential anti-facts result of each user-article pair based on the average causal effect; Fusing the potential anti-facts results with the original records in the observation data set to construct a training data set containing anti-facts samples; Training the depth recommendation model based on the training data set to obtain a target depth recommendation model; And predicting candidate articles according to the target depth recommendation model aiming at target users to obtain target predicted values, and generating a recommendation list according to the target predicted values.
  2. 2. The causal inference based anti-facts recommendation method of claim 1, wherein said employing a dual machine learning mechanism to estimate an average causal effect of exposure behavior on user feedback based on said observation dataset and said trend score, and to calculate potential anti-facts results for each user-item pair based on said average causal effect, comprises: Training a first machine learning model based on samples in the observed dataset for which exposure indications are true to fit a mapping of the user features and the item features to user feedback variables; acquiring a predicted value of an exposure indicating variable based on the trend score; Respectively calculating a first residual error between the observed value of the user feedback variable and a predicted value of a first machine learning model and a second residual error between the observed value of the exposure indicating variable and the predicted value; estimating the average causal effect of the exposure behavior on user feedback based on the first residual error and the second residual error by adopting a linear regression mode; And calculating the potential anti-facts result of each user-object pair in the two states of the exposure behavior and the non-exposure behavior based on the average causal effect and the predicted value of the first machine learning model.
  3. 3. The causal inference based anti-facts recommendation method of claim 1, wherein the fusing of the potential anti-facts results with the original records in the observation dataset to construct a training dataset comprising anti-facts samples comprises: constructing an observation sample subset by indicating exposure in the observation data set as a true sample and true feedback; the potential anti-facts results, the unexposed potential results corresponding to the observed samples in the observed dataset and the exposed potential results corresponding to the unobserved samples are combined to form an anti-facts sample subset; According to the original record in the observation data set and the tendency score, distributing a calibration weight to each sample in the anti-facts sample subset to obtain a weighted anti-facts sample subset; And combining the observation sample subset with the weighted anti-facts sample subset to generate the training data set.
  4. 4. The causal inference based anti-facts recommendation method of claim 1, wherein said obtaining a recommendation system generated observation data set, constructing a causal graph based on potential relationships between the recommendation system's business logic and variables in the observation data set, comprises: obtaining an observation data set generated by the recommendation system, wherein the observation data set comprises a user characteristic set, an article characteristic set, an exposure strategy variable, an exposure indicating variable and a user feedback variable; And constructing a causal link and an action relation between each node based on the observation data set and service logic of the recommendation system to construct the potential relation and construct the causal graph, wherein the causal graph defines the causal direction and the action relation between the nodes, and the causal link comprises a direct causal link and an indirect causal link.
  5. 5. The causal inference-based anti-facts recommendation method of claim 1, wherein the training depth recommendation model based on the training data set, to obtain a target depth recommendation model, comprises: encoding user features and object features in the training data set into feature vectors with unified dimensions; calculating attention weight according to the tendency score, and carrying out weighted adjustment on the feature vector based on the attention weight to obtain an adjusted feature vector; outputting feedback prediction probability of the user on the article through the adjusted feature vector, and calculating a model loss value based on the feedback prediction probability; And adjusting model parameters of the depth recommendation model through the model loss value to carry out model iterative training again, so as to obtain the target depth recommendation model.
  6. 6. The causal inference-based anti-facts recommendation method according to any one of claims 1 to 5, wherein said predicting candidate items for a target user by the target depth recommendation model to obtain a target predicted value, and generating a recommendation list according to the target predicted value, comprises: acquiring user characteristics corresponding to the candidate items and the target user; extracting the characteristics of each candidate object to obtain the characteristics of the target object; Predicting the target depth recommendation model based on the user characteristics and the target object characteristics to obtain the target predicted value; and sequencing all the target predicted values to obtain a sequencing result, and generating the recommendation list based on the sequencing result.
  7. 7. The causal inference based anti-facts recommendation method according to any of claims 1 to 5, wherein the classification model is a logistic regression model, the training objective of which is to minimize the prediction error of exposure indicating variables, and wherein model parameters of the logistic regression model are used to quantify the causal impact weights of each dimension of the user profile and the item profile on the exposure probability.
  8. 8. A causal inference-based anti-facts recommender comprising: the causal graph construction module is used for acquiring an observation data set generated by a recommendation system and constructing a causal graph based on potential relations between service logic of the recommendation system and variables in the observation data set; A trend score estimation module for estimating a trend score by a classification model based on causal relationships in the causal graph, taking as input user features and item features in the observation dataset; a counterfactual result calculation module for estimating an average causal effect of exposure behavior on user feedback based on the observation dataset and the trend score using a dual machine learning mechanism, and calculating potential counterfactual results for each user-item pair based on the average causal effect; The training data set construction module is used for fusing the potential anti-facts results with the original records in the observation data set to construct a training data set containing anti-facts samples; The recommendation module training module is used for training the depth recommendation model based on the training data set to obtain a target depth recommendation model; And the recommendation list generation module is used for predicting candidate articles according to the target depth recommendation model aiming at target users to obtain target predicted values, and generating a recommendation list according to the target predicted values.
  9. 9. A computer device comprising a memory and a processor, the memory having stored therein a computer program, the processor implementing the causal inference based anti-facts recommendation method of any of claims 1 to 7 when the computer program is executed.
  10. 10. A computer readable storage medium, having stored thereon a computer program which, when executed by a processor, implements the causal inference based anti-facts recommendation method of any of claims 1 to 7.

Description

Anti-facts recommendation method, device, equipment and storage medium based on causal inference Technical Field The application relates to the technical field of artificial intelligence, and can be applied to the medical field and the financial science and technology field, in particular to a causal inference-based anti-facts recommendation method, device, equipment and storage medium. Background Along with the wide application of big data and artificial intelligence technology, personalized recommendation systems play a key role in e-commerce, content distribution, financial services, medical health and other scenes. The traditional recommendation method mainly relies on statistical association in historical interaction data to predict, for example, a collaborative filtering algorithm is used for mining similarity through a user-object interaction matrix, content-based recommendation is used for matching user preference through object attribute characteristics, and a deep learning model is used for capturing complex modes through a neural network. However, such correlation-based predictive paradigms suffer from inherent drawbacks, particularly in cases where data is subject to systematic deviations, the recommendation and fairness are significantly degraded. In the field of financial science and technology, such as bank financial management or insurance product recommendation, historical data deviation can ignore real financial conditions and risk bearing capacity of a user, cause incorrect product matching, cause customer loss or investment hidden danger, and meanwhile, high-quality mass products are difficult to reach target mass, in medical health scenes, such as clinical auxiliary decision making or health management scheme recommendation, historical data only comprise treatment schemes which have been prescribed by doctors, strong selection deviation exists, and if the system is directly trained based on the data, the existing diagnosis and treatment path can be solidified, exploration of potential better schemes is inhibited, and personalized treatment and medical resource optimization are not facilitated. The existing causal inference theory provides a new thought for breaking through the limitation, and is characterized in that causal relationships and statistical correlations are distinguished, the anti-facts inference is used as a key tool, and the results of 'if different actions are taken' can be simulated, so that a user decision process is modeled from a causal layer, observation deviation is corrected, and the real interest of a user in an unexposed item is estimated. While the prior art attempts to introduce causal concepts, the lack of a systematic framework to integrate causal graph construction, trend score estimation and anti-facts result calculation makes it difficult to effectively process high-dimensional features and complex interactions in a recommendation scene, and an end-to-end method is needed to solve the problem of data deviation and achieve fairer, more accurate and interpretable recommendation. Disclosure of Invention The embodiment of the application aims to provide a causal inference-based anti-facts recommendation method, device, equipment and storage medium, which can effectively correct exposure deviation, selection deviation and confounding factor interference in observed data, so as to provide a fair, accurate and interpretable recommendation result. In order to solve the above technical problems, an embodiment of the present application provides a counterfacts recommendation method based on causal inference, including: Acquiring an observation data set generated by a recommendation system, and constructing a causal graph based on potential relations between service logic of the recommendation system and variables in the observation data set; Based on the causal relationship in the causal graph, taking the user features and the object features in the observation data set as input, and estimating a tendency score through a classification model; estimating an average causal effect of exposure behavior on user feedback based on the observation data set and the trend score by adopting a double machine learning mechanism, and calculating a potential anti-facts result of each user-article pair based on the average causal effect; Fusing the potential anti-facts results with the original records in the observation data set to construct a training data set containing anti-facts samples; Training the depth recommendation model based on the training data set to obtain a target depth recommendation model; And predicting candidate articles according to the target depth recommendation model aiming at target users to obtain target predicted values, and generating a recommendation list according to the target predicted values. In order to solve the above technical problems, an embodiment of the present application provides a counterfacts recommendation device based on causal infere