A recommendation system based on feature weight model and an interpretability method of recommendation results

Abstract

The present invention relates to a method for interpretability of recommendation results of a feature weight model recommendation system. The method integrates multi-way recall strategy names, sorting model TopK weight features and post-processing rule names to achieve interpretability of recommendation results of the recommendation system. When a recommendation request is generated, the recommendation system will process the materials to be recommended through recall, sorting and post-processing processes, obtain the recommendation results and return them to the requester. It also relates to a recommendation system based on a feature weight model. The method and system of the present invention can effectively explain the reasons for the existence of material ID in the recommendation results and the reasons for the relative position order, avoiding the unexplainable nature of the recommendation results being completely black box, and improving the interpretability of the recommendation system from this perspective.

Description

Technical Field

[0001] The present invention relates to artificial intelligence, and in particular to a method for interpreting recommendation results of a feature weight model recommendation system. Background Art

[0002] The recommendation system is becoming more and more integrated into people's daily lives, connecting thousands of material resource data to a wide range of users, providing each person with an exclusive recommendation result space, and achieving personalized recommendation effects for each person.

[0003] Recommendation systems generally include several main processes or modules such as recall, sorting and post-processing. The recommendation results are often the result of the combined effect of these processes or modules. However, in some fields or industries, clear and detailed requirements are increasingly being put forward for the interpretability of recommendation results. This poses new challenges for us to deeply understand the scope and input and output of the internal processes or models of the recommendation system.

[0004] Chinese invention patent 202110151400.2 discloses an explainable recommendation system based on knowledge graph, which provides high prediction accuracy and improves the purpose of very efficient interpretation, and has high practical value and promotion value. However, this aspect adopts a recommendable system based on knowledge graph, which belongs to a recommendation interpretation scheme situation and has a limited scope of application. In addition, the recall process of the existing recommendation system can be summarized as multi-way (model or strategy) recall and single deep model recall. The sorting process is generally an LTR (learning to rank) sorting model, plus post-processing rules to form a recommendation engine, which can often provide better recommendation effects and recommendation indicators, but the work on the interpretability of the recommendation results has made limited progress. The lack of work in this area may bring the following adverse effects: 1. It is impossible to explain the rationality of the recommendation results to the outside world. The recommendation system is still equivalent to a black box, which cannot well reflect the intrinsic value of the recommendation system and gain wide recognition, which undermines the depth and quality of the work report. 2. It is not conducive to the optimization of recommendation effects and the recommendation system. Good interpretability of recommendation results can help us improve the recommendation system, but otherwise it is not conducive to improving the recommendation system, making the upgrade and transformation of the recommendation system lose direction. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a recommendation result explainability method based on a feature weight model recommendation system. The recommendation result explainability method of the present invention can accurately recommend request results and effectively explain the reasons why material IDs exist in the recommendation results and the reasons for the relative position order.

[0006] In order to achieve the above invention purpose, the technical solution provided by the present invention is as follows:

[0007] A method for interpretability of recommendation results of a feature weight model recommendation system is proposed. The method integrates the multi-way recall strategy name, the TopK weight feature of the sorting model and the post-processing rule name to achieve the interpretability of the recommendation results of the recommendation system, and effectively explains the reasons why the material ID exists in the recommendation results and the reasons for the relative position order. The method specifically includes the following implementation steps:

[0008] The first step is the recommendation request. When a recommendation request is generated, the recommendation system will process the recommended materials through subsequent recall, sorting and post-processing processes;

[0009] The second step is the recall process, obtaining the recall strategy name for multi-way recall;

[0010] The third step is the sorting process, which obtains the TopK weight features of the sorting model, records the feature values ​​during a single sorting prediction, and selects the key features of this sorting prediction;

[0011] The fourth step is post-processing, which is to obtain the names of various rules that act on the recommendation results in the post-processing process;

[0012] Step 5: Obtain recommendation results and explain the recommendation results, which is completed based on the recall strategy name, sorting key features and post-processing rule name;

[0013] The sixth step is to complete the interpretable presentation of the recommendation results and return the obtained recommendation results to the requester.

[0014] In the recommendation result explainability method based on the feature weight model recommendation system of the present invention, the recommendation system adopts a multi-way recall process in the second step to record the recall strategy name recall_name, recall strategy weight recall_weight and relative position order recall_position generated by each material ID in the recall result set.

[0015] In the recommendation result explainability method based on the feature weight model recommendation system of the present invention, the third step passes the recall result into the sorting process, and predicts the click probability of each material ID of the recall result through the sorting model, which is used as the basis for re-sorting to obtain the sorting result. This process involves model selection, feature engineering, model training and model prediction process of the sorting model.

[0016] In the method for interpretability of recommendation results based on the feature weight model recommendation system of the present invention, further, in the selection of the ranking model, a ranking model Model that can reflect the feature weight is selected, such as the LR model; the LR model formula is:

[0017]

[0018] In the above formula, x represents the feature Feature, xi represents a certain feature value, and wi represents the feature weight corresponding to the feature. After normalizing all feature values ​​feature_value, the feature value domain space is compressed to [0,1], and the LR model is trained to obtain the feature weights w corresponding to all feature Features. The higher the weight, the greater the impact on model reasoning.

[0019] In the method for interpretability of recommendation results based on the feature weight model recommendation system of the present invention, further, in feature engineering, feature selection, feature preprocessing, feature construction, feature intersection, feature post-processing and other links are performed to obtain the feature Feature and feature value feature_value to be finally input into the sorting model.

[0020] In the recommendation result explainability method based on the feature weight model recommendation system of the present invention, further, in the model training, the feature Feature is input into the sorting model Model. After the LR model training, the relative weight of each input feature can be obtained. These feature weights reflect the importance of the feature Feature to the model reasoning. Here, the feature Feature@Top50% of the top 50% of the high-weight features is obtained.

[0021] In the interpretability method of recommendation results based on the feature weight model recommendation system of the present invention, further, in model prediction, according to the feature Feature determined during model training, the predicted feature value feature_predict corresponding to each material ID in the recall result set is input to obtain the predicted click probability click_probility of the recalled material ID. At this time, the feature Feature_predcit with a value in feature_predict is taken to intersect with Feature@Top50%, and the real-time high-weight feature Feature_predict@Top50% when predicting the click rate of each material ID is obtained, and the ranking position order rank_position after prediction by the ranking model.

[0022] In the method for interpretability of recommendation results based on the feature weight model recommendation system of the present invention, in the fourth step, the sorting results are passed to the post-processing process, and the final recommendation results are obtained through the post-processing process, and the post-processing rule name postprocess_name applied to the recommendation results and the final position order of the material ID in the recommendation results are recorded. Post-processing is a conventional operation and is a simple rule selection.

[0023] In the recommendation result explainability method based on the feature weight model recommendation system of the present invention, in the fifth step, the recommendation reasons generated by each material ID in the final recommendation result are obtained, and the reasons for the existence and relative position of the recommendation result material ID are explained. The recommendation reasons are composed of the following: recall_name+recall_weight+recall_position+Feature_predict@Top50%+rank_position+postprocess_name+final_position.

[0024] Based on the above technical solution, the recommendation result explainability method based on the feature weight model recommendation system of the present invention has the following technical advantages compared with the prior art:

[0025] 1. The present invention realizes the interpretability of the recommendation results of the recommendation system based on the feature weight model by integrating the multi-way recall strategy name, the TopK weight feature of the sorting model and the post-processing rule name, and effectively explains the reasons why the material ID exists in the recommendation results and the reasons for the relative position order. The TopK weight feature refers to the weight feature of the first K percent.

[0026] 2. In the explainability method of recommendation results of the present invention, the recall results are passed into the sorting process, and the click probability of each material ID in the recall results is predicted by the sorting model, which is used as the basis for re-sorting to obtain the sorting results. This process involves model selection, feature engineering, model training and model prediction processes of the sorting model, so as to obtain accurate recommendation results and give clear explanations for the recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a schematic diagram of the operation flow of a method for interpreting recommendation results of a feature weight model recommendation system according to the present invention. DETAILED DESCRIPTION

[0028] Below we further explain in detail the recommendation result explainability method based on the feature weight model recommendation system of the present invention in conjunction with the accompanying drawings and specific embodiments, in order to more clearly understand its step composition and operation mode, but this cannot be used to limit the protection scope of the present invention.

[0029] like Figure 1As shown, the present invention is a method for interpreting the recommendation results of a feature weight model recommendation system. In general, the method for interpreting the recommendation results of the present invention integrates the multi-way recall strategy name, the TopK weight feature of the sorting model and the post-processing rule name to achieve the interpretability of the recommendation results of the recommendation system, and effectively explains the reasons why the material ID exists in the recommendation results and the reasons for the relative position order. Specifically, the method includes the implementation steps:

[0030] The first step is the recommendation request. When a recommendation request is generated, the recommendation system will process the recommended materials through subsequent recall, sorting and post-processing processes;

[0031] The second step is the recall process, obtaining the recall strategy name for multi-way recall;

[0032] The third step is the sorting process, which obtains the TopK weight features of the sorting model, records the feature values ​​during a single sorting prediction, and selects the key features of this sorting prediction;

[0033] The fourth step is post-processing, which is to obtain the names of various rules that act on the recommendation results in the post-processing process;

[0034] Step 5: Obtain recommendation results and explain the recommendation results, which is completed based on the recall strategy name, sorting key features and post-processing rule name;

[0035] The sixth step is to complete the interpretable presentation of the recommendation results and return the obtained recommendation results to the requester.

[0036] In the recommendation result explainability method based on the feature weight model recommendation system of the present invention, the recommendation system adopts a multi-way recall process in the second step to record the recall strategy name recall_name, recall strategy weight recall_weight and relative position order recall_position generated by each material ID in the recall result set.

[0037] In the recommendation result explainability method based on the feature weight model recommendation system of the present invention, the third step passes the recall result into the sorting process, and predicts the click probability of each material ID of the recall result through the sorting model, which is used as the basis for re-sorting to obtain the sorting result. This process involves model selection, feature engineering, model training and model prediction process of the sorting model.

[0038] In the method for interpretability of recommendation results based on the feature weight model recommendation system of the present invention, further, in the selection of the ranking model, a ranking model Model that can reflect the feature weight is selected, such as the LR model; the LR model formula is:

[0039]

[0040] In the above formula, x represents the feature Feature, xi represents a certain feature value, and wi represents the feature weight corresponding to the feature. After normalizing all feature values ​​feature_value, the feature value domain space is compressed to [0,1], and the LR model is trained to obtain the feature weights w corresponding to all feature Features. The higher the weight, the greater the impact on model reasoning.

[0041] In the method for interpretability of recommendation results based on the feature weight model recommendation system of the present invention, further, in feature engineering, feature selection, feature preprocessing, feature construction, feature intersection, feature post-processing and other links are performed to obtain the feature Feature and feature value feature_value to be finally input into the sorting model.

[0042] In the recommendation result explainability method based on the feature weight model recommendation system of the present invention, further, in the model training, the feature Feature is input into the sorting model Model. After the LR model training, the relative weight of each input feature can be obtained. These feature weights reflect the importance of the feature Feature to the model reasoning. Here, the feature Feature@Top50% of the top 50% of the high-weight features is obtained.

[0043] In the interpretability method of recommendation results based on the feature weight model recommendation system of the present invention, further, in model prediction, according to the feature Feature determined during model training, the predicted feature value feature_predict corresponding to each material ID in the recall result set is input to obtain the predicted click probability click_probility of the recalled material ID. At this time, the feature Feature_predcit with a value in feature_predict is taken to intersect with Feature@Top50%, and the real-time high-weight feature Feature_predict@Top50% when predicting the click rate of each material ID is obtained, and the ranking position order rank_position after prediction by the ranking model.

[0044] In the interpretability method of recommendation results based on the feature weight model recommendation system of the present invention, the sorting results are passed to the post-processing process in the fourth step, and the final recommendation results are obtained through the post-processing process, and the post-processing rule name postprocess_name applied to the recommendation results and the final position order final_position of the material ID in the recommendation results are recorded.

[0045] In the method for interpretability of recommendation results based on the feature weight model recommendation system of the present invention, in the fifth step, the recommendation reasons for each material ID in the final recommendation result are obtained, and the reasons for the existence and relative position of the material ID in the recommendation result are explained. The recommendation reasons are composed of the following:

[0046] Recall strategy name recall_name + recall strategy weight recall_weight + relative position order recall_position + high-weight features Top 50% features Feature_predict@Top50% + ranking position order rank_position + post-processing rule name postprocess_name + position order final_position.

[0047] Undoubtedly, the above is only one implementation process of the patent application of the present invention, and other similar operation processes and implementation methods are also included. In short, the protection scope of the present invention also includes other obvious changes and substitutions for those skilled in the art.

Claims

1. A method for interpretability of recommendation results of a feature weight model recommendation system, characterized in that: This method integrates the names of multi-way recall strategies, the TopK weight features of the sorting model, and the names of post-processing rules to achieve the interpretability of the recommendation results of the recommendation system, and effectively explains the reasons why the material ID exists in the recommendation results and the reasons for the relative position order. The method specifically includes the following implementation steps: The first step is the recommendation request. When a recommendation request is generated, the recommendation system processes the materials to be recommended through subsequent recall, sorting and post-processing processes; The second step is the recall process, obtaining the recall strategy name for multi-way recall; The third step is the sorting process. The TopK weight features of the sorting model are obtained, and the feature values ​​of a single sorting prediction are recorded. The key features of this sorting prediction are screened out. The click probability of each material ID in the recall result is predicted by the sorting model, which is used as the basis for re-sorting to obtain the sorting result. This process involves the model selection, feature engineering, model training and model prediction process of the sorting model. In the model prediction, according to the feature Feature determined during model training, the predicted feature value feature_predict corresponding to each material ID in the recall result set is input to obtain the predicted click probability click_probility of the recalled material ID. The feature Feature_predcit with a value in the predicted feature value feature_predict is taken to intersect with Feature@Top50%. The feature Feature_predcit with a value is a true subset of the predicted feature value feature_predict. The real-time high-weight feature Feature_predict@Top50% when predicting the click rate of each material ID is obtained, and then the sorting position order rank_position after the sorting model prediction is obtained. The fourth step is post-processing, which obtains the names of various rules that act on the recommendation results in the post-processing process; Step 5: Get the recommendation results, explain the recommendation results, complete the explanation based on the recall strategy name, ranking key features and post-processing rule name, obtain the recommendation reasons for each material ID in the final recommendation result, explain the reasons for the existence of the recommended result material ID and the relative position reasons. The recommendation reason consists of the following: recall strategy name recall_name + recall strategy weight recall_weight + relative position order recall_position + real-time high weight feature Feature_predict@Top50% + ranking position order rank_position + post-processing rule name postprocess_name + final position order final_position in the recommendation result; The sixth step is to complete the interpretable presentation of the recommendation results and return the obtained recommendation results to the requester.

2. The method for interpretability of recommendation results based on a feature weight model recommendation system according to claim 1, characterized in that: In the second step, the recommendation system adopts a multi-way recall process, and records the recall strategy name recall_name, recall strategy weight recall_weight and relative position order recall_position generated by each material ID in the recall result set.

3. The method for interpretability of recommendation results based on a feature weight model recommendation system according to claim 1, characterized in that: In the selection of the sorting model, a sorting model that can reflect the feature weight is selected, including an LR model. The LR model formula is: ， In the above formula, xi represents a certain eigenvalue, wi represents the feature weight corresponding to the feature, and after normalizing all eigenvalues ​​feature_value, the eigenvalue domain space is compressed to [0, 1], and the LR model is trained to obtain the feature weights corresponding to all features Feature.

4. The method for interpretability of recommendation results based on a feature weight model recommendation system according to claim 3, characterized in that: In feature engineering, feature selection, feature preprocessing, feature construction, feature intersection and feature post-processing are performed to obtain the feature Feature and feature value feature_value to be input into the final sorting model.

5. The method for interpretability of recommendation results based on a feature weight model recommendation system according to claim 3, characterized in that: When training the LR model, the feature Feature is input into the sorting model. After the LR model is trained, the relative weight of each input feature is obtained. The feature weight reflects the importance of the feature Feature to the model reasoning. The features of the top 50% of high-weight features are obtained and expressed as Feature@Top50%.

6. The method for interpretability of recommendation results based on a feature weight model recommendation system according to claim 1, characterized in that: In the fourth step, the sorting results are passed to the post-processing process, and the final recommendation results are obtained through the post-processing process. The post-processing rule name postprocess_name applied to the recommendation results and the final position order of the material ID in the recommendation results final_position are recorded.

7. A method for interpreting recommendation results of a feature weight model recommendation system according to any one of claims 1 to 6, characterized in that: This method uses a feature weight model-based recommendation system, which integrates multi-way recall strategy names, sorting model TopK weight features and post-processing rule names to achieve the interpretability of recommendation results. Its composition includes a recall process module, a sorting process module and a post-processing process module. The recall process module obtains the recall strategy name of the multi-way recall; The process module obtains the TopK weight features of the sorting model, records the feature values ​​during a single sorting prediction, and screens out the key features of this sorting prediction; the post-processing process module obtains the names of various rules that act on the recommendation results in the post-processing process; when a recommendation request is generated, the recommendation system processes the materials to be recommended through subsequent recall, sorting and post-processing processes, obtains the recommendation results, and explains the recommendation results.

Images