Research On Robust Recommendation Algorithm Based On Attack Detection And Matrix Decomposition

In the current era of big data,search engines and recommendation technologies are generally adopted to deal with the problem of massive data,among which recommendation technologies can achieve personalized and proactive recommendation effects.However,with the development of the recommendation technology,Attack(Shilling Attack)also arise,supporting some characteristics of the Attack using the recommended technology own existence,disguised as a normal user,generate false information user profile.Malicious users can achieve personal malicious purposes by injecting a certain amount of user profile information,changing the recommendation results,and affecting the recommendation performance of the recommendation system.There are two main ways to deal with the toto attack: one is to filter the attack through attack detection before recommendation.The other is to improve the robustness of the recommendation algorithm and reduce the impact of the toto attack.There are many ways to improve the recommendation performance of the recommendation algorithm,such as excellent new recommendation algorithm,improvement of the original algorithm,multiple iterations,and fusion of multiple recommendation algorithms.In dealing with toto attacks,this paper adopts the first scheme of attack detection,filtering attacks before recommendation.In the aspect of recommendation accuracy,this paper,by using the matrix decomposition using lingo righteousness Model(Latent Factor Model),continuously improve recommendation performance optimization parameters.In order to improve the robustness of the system under the premise of ensuring the recommendation precision,this paper proposes a robust algorithm.Attack detection ensures robustness and matrix decomposition recommendation algorithm ensures recommendation accuracy.Specific research contents are as follows:(1)for attack detection,based on the SVM-KNN a semi-supervised attack detection classifier,combined Support Vector Machine(Support Vector Machine,abbreviated as SVM)and K nearest neighbor(K-NearestNeighbor,abbreviated as KNN)to test the attack.Firstly,SVM is classified as a basic classifier,and KNN classifier is introduced to classify the points near the boundary surface to improve the detection accuracy in view of the poor classification effect of attachment points on the boundary surface of SVM.For KNN,the similarity calculation formula is improved and the detection performance is improved.Finally,the idea of semi-supervision is used to improve the detection accuracy by using both marked data and unmarked data.,(2)for recommendation algorithm was proposed based on the recommendation of matrix decomposition algorithm,using lingo righteousness model,the optimization of parameters,and through constant iterative stochastic gradient descent method,further improve the recommendation accuracy.(3)the attack detection algorithm and matrix decomposition recommendation algorithm are used together.First,the attack detection algorithm is used to filter out the truncation attack,and then the matrix decomposition recommendation algorithm is used for recommendation.Simulation experiments were carried out on the Movielens data set,which achieved the desired goal and improved the robustness of the system with the guarantee of recommended performance.