Recommendation Algorithms Research Of Latent Factor And Multi-attribute Content Filtering

In the actual system, the information keep on increasing explosively, users are eger tofind the interest items in a lot of information, building recommendation systems can helpusers to solve this problem. The sparsity problem of recommendation refers to the majority ofusers only buy a few items. The cold start problem refers to a new user or a new item has nothistory, and the recommendation system can’t recommend. Matrix factorization and sparselinear model can factorize score matrix to potential latent matrixs for user and item. For thedata sparsity problem, study the matrix factorization and sparse linear model. We proposeTrustSLIM, a trust-based sparse linar model. Do a detailed derivation and experimentalverification for the solution algorthim. As to the cold start problem, proposing the model thatmulti-attribute content filtering. Do research and improvement of combing different attributes,specific contents are as follows:(1) Research the matrix factorization model, combined with ratings of user history andthe context of user trust, using the gradient descent method avoids the problem of datasparsity.(2) Research the sparse linear model, this particular matrix decomposition model, theinterpretation and complexity is superior to matrix factorization model. Using the coordinatedescent method solve the optimizing problem that containing a norm optimization function.We introduce trust factor and build the TrustSLIM model. The model can effectivelyrecommend topN items. Experiments certificated that accuracy and recall are improvedeffectively.(3) Deeply research the multi-attribute content filtering model. Firstly defined theprobability score of age for the item, on behalf of the degree that the the user interest for the item. Proposing the method of comparing the error with the standard actural score select theattribute to recommend score. This method solves the problem of combing different attributes.We also fit the probability scores using support vector regression, and the experimentsvertificate that the methods are effective.(4) Discussed mixed stragety between potential factor and multi-attribute content filteringrecommendation algorithm...