Feature Recommendation Method and Apparatus

Abstract

The present invention provides a feature recommendation method and device. The feature recommendation method includes: determining a target value of text features in sample data according to an output text feature estimation model, wherein the output text feature estimation model is selected from training data according to The optimal combination of features is obtained; the text features in the sample data are sorted according to the target value, and the text features in the sample data are recommended according to the target value in descending order. The invention can realize automatic selection of effective combined features, save time and effort, effectively solve the problem of time-consuming and labor-intensive in the existing manual feature selection process, and can improve the effectiveness of the recommendation system.

Description

technical field [0001] The present invention relates to the technical field of the Internet, in particular to a feature recommendation method and device. Background technique [0002] In the prior art, the text recommendation system usually adopts the following methods when selecting features: [0003] 1. Selection by Factorization Machines (hereinafter referred to as: FM), where FM is a generalized model, mainly used to model all pairwise interaction features, and the parameters of the interaction features are passed through the shared low-rank vector inner product get; [0004] 2. Select by random split tree algorithm, specifically, use text information to separate the user item matrix into sub-matrices according to specific text values, and then perform matrix decomposition for each sub-matrix, and the final predicted value is the average of T generated decision tree predictions value. [0005] However, all pairwise interaction features are simulated in FM, but no effe...

AI Technical Summary

Problems solved by technology

[0005] However, all pairwise interaction features are simulated in FM, but no effective feature combination is selected. In reality, some interaction features may be invalid. In the FM model, the weights of all interaction features are shared through the low-rank Vector inner product to obtain, if an interaction feature is invalid, it will lead to inaccurate parameter estimation and final result prediction

[0006] In addition, the random split tree algorithm does not discuss the selection of combined features. When there are dozens of discrete features, the random split tree algorithm is not very effective.

[0007] To sum up, the existing technology has the problem that the combined features cannot be effectively selected, and the effectiveness of the recommended text features is low.

Images