Recommendation Algorithms Based On Combination Of Matrix Factorization And Random Walk

In today’s era of information explosion, people can find quickly the information content of interest with the help of recommendation systems which are powerful information filtering tools. Collaborative filtering is the most popular category of algorithms in recommendation systems. Matrix factorization and random walk are two different kinds of model-based collaborative filtering algorithms applied in recommendation systems.The matrix factorization model can deal with the sparsity of training data by mapping users and items into the latent factor space and thus has good scalability. Factorization machines (FM) are a generic approach and can mimic many kinds of matrix factorization models just by feature engineering. To facilitate the use of various matrix factorization models, we implemented the FM and made some optimizations such as user-based feature grouping and caching.Random walk is another method which can be used to alleviate the sparsity problem. There are multiple versions of random walk implementations whose transitive probabilities are all computed heuristically. Here, a model-based method is proposed to obtain the transitive probability matrix. The method maps the nodes in the random walk model into the latent factor vector through matrix factorization and computes the transitive probabilities between nodes using the vector. The validity of the method was demonstrated on a specific random walk model. The experimental results show that the method can improve the prediction performance of random walk model with better stability. It can effectively capture the transitive probabilities between nodes even for extremely sparse datasets.