Research For Collaborative Filtering Algorithm Regarding Sparsity

Collaborative filtering(CF)is the most popular and successful recommendation algorithm in recommendation system.Based on the idea of collective intelligence,CF selects neighbors whose behavior resemble target ser and recommends product according to neighbors'interests.Although CF can solve 'information overload',it still faces various challenges.Data sparsity is the main issue blocking its progress.Due to the excessive sparsity of rating data,similarity calculation and neighbor selection could be inaccurate,affecting the accuracy and reliability of recommendation results.This article performed an in-depth research regarding data sparsity issue and introduced the following three optimized collaborative filtering algorithms.1.Optimized algorithm based on filling predicted value and user's/item's mean absolute error(MAE).This algorithm calculates user's/item's MAE and fills the empty values in the rating matrix with the predicted value and MAE under the filling rule.And then uses the filled rating matrix to perform CF algorithm.By ensuring the filling values near to users'average value,the algorithm not only guarantees the predicted results to satisfy users'personalized rating habits,but also improves recommendation's accuracy.2.Optimized algorithm based on item-clustering.This algorithm clusters the item list of the original matrix,structures two 'user-category' matrix with higher density and calculates two kinds of similarities accordingly.These similarities are linear weighted and combined with correction factor as the final similarity to perform recommendation.Calculating similarities based on a denser matrix leads to a more reliable recommendation.Meanwhile,algorithm efficiency is also improved due to reduced scale of matrix.3.Optimized algorithm based on trust network.This algorithm introduces trust relations into collaborative filtering.By using co-ratings among users and propagation rules to build a trust network.unrated items are predicted through linear weighing the trust degree of trust network and traditional user similarity.This algorithm recommends according to the preferences of target users' trusts,making the accuracy higher and the result more reliable than what based on neighbors.A large amount of comparative experiments demonstrated that algorithms showed in this article are effective and feasible.