Semi-supervised learning-based recommendation system shilling attack detection method

Abstract

The invention discloses a semi-supervised learning-based recommendation system shilling attack detection method. The method comprises the following stages: a shilling attack detection preprocessing stage, namely preprocessing the data of a marked data set and an unmarked data set to obtain shilling attack detection indexes of the marked and unmarked data sets, and training an initial Naive Bayes classifier on the marked data set; and an EM-lambda algorithm stage, namely continuously iterating by an EM-lambda algorithm to obtain a stable classifier and finally obtaining the type of the unmarked data set, and predicting whether an unmarked user belongs to a normal user (N) or a shilling user (S) through a function f: U to C so as to finish the recommendation system shilling attack detection. The method is used for discovering shilling attack users in a recommendation system, has high efficiency, sensitivity, special effect performance, high detection rate and low error rate.

Description

technical field [0001] The invention belongs to the field of information security, in particular to a semi-supervised learning-based recommendation system trolling attack detection method. Background technique [0002] The rapid growth of the scale and coverage of the Internet has brought about information overload. We live in the data age. According to IDC estimates, the total amount of data will reach 1.8ZB (1ZB=1,000,000PB) in 2011. Users are helpless in the face of massive data, and it is difficult to find useful information from it. Information. The recommender system is an important means of information filtering and a very potential method to solve the problem of information overload. The recommendation algorithm is the core part of the entire recommendation system. Collaborative filtering is the most widely used recommendation algorithm. In daily life, we often use the recommendations of good friends to make some choices. Collaborative filtering is based on this ide...

AI Technical Summary

Problems solved by technology

[0005] In the actual recommendation system, there are often a large number of users whose identities cannot be determined (called unlabeled data), while only a small number of users can be identified (called labeled data), such as users with extremely high or low praise rates on Taobao, It is easy to determine the identity of Huangguan users, etc., but it is difficult to determine the identity of a large number of moderately favorable users

Due to the very small amount of labeled data, the error of the detection method based on supervised learning is relatively large

On the other hand, unsupervised learning does not use the existing labeled data set, and ignores the influence of the labeled data set on the unlabeled data set. The results of training are often difficult to have strong generalization learning ability.

Images