Score prediction method combining topic model and heterogeneous information network

Abstract

The invention discloses a score prediction method combining a topic model and a heterogeneous information network. The method comprises the following steps: (1) extracting comment information by usinga topic model for a specified user commodity pair, thereby constructing vector representation of a user and a commodity; (2) constructing a heterogeneous information network by using the commodity attribute information and the user common purchase information; (3) extracting a final relationship representation vector of the user commodity pair from the heterogeneous network; (4) for the user commodity pair, connecting the user vector, the relationship representation vector and the commodity vector representation, and inputting the user vector, the relationship representation vector and the commodity vector representation into the AFM to realize score prediction; and step 5, calculating an RMSE value according to the prediction scoring data and the real scoring data calculated by the model, and taking the RMSE value as an evaluation index of the model effect. According to the method, the problems of cold start, low interpretability and low accuracy in scoring prediction are solved.

Description

technical field [0001] The invention relates to the field of score prediction in recommendation systems, in particular to a score prediction method for joint topic models and heterogeneous information networks. Background technique [0002] There are currently three types of methods for predicting users' ratings for unpurchased products: [0003] The first is to use collaborative filtering to achieve rating prediction based on user historical rating records. This method will cause cold start problem for sparse user rating data. [0004] The second is to use comment and rating information to achieve rating prediction. LDA is a relatively traditional method of analyzing comment data, but currently more methods are used to process comment information through deep learning such as CNN. This method ignores the inherent attribute information of the product and the prediction effect of the user's common purchase information is not good. [0005] The third is to construct a heter...

AI Technical Summary

Problems solved by technology

[0006] Aiming at the deficiencies of the prior art, the present invention provides a score prediction method that combines topic models and heterogeneous information networks to solve the problems of cold start, poor interpretability, and low accuracy in score prediction

Images