Research On Collaborative Filtering Algorithm Based On Preference Of Item Attributes And Social Tags

Recently,online shopping,e-commerce and online sales become more and more popular.According to the authoritative statistics of China Internet Network Information Center in 2017,only the user scale of mobile online payment is up to 470 million in China.Therefore,it's important for customers to achieve accurate recommendations.And it's necessary to provide users with the most valuable goods for all businesses.However,the personalized recommended system is the best strategy to solve this problem.Moreover,collaborative filtering algorithm is the most widely-used and successful recommendation algorithm in recommendation system,but there still exist several drawbacks that affect the recommended quality.The traditional collaborative filtering algorithm has the following drawbacks: The first one is the problem of data sparsity.The user's rating matrix is too sparse to measure the similarity of users or items effectively.The second one is the change of the user's interest.As the user's interest maybe change over time,thus it is unreasonable and can not reflect the dynamic changes on users' interest if considering all the time of rating are the same weight.The last problem is Scalability.The scale of the score matrix is too high and may resulting in complex computational cost with the increasing amount of data.Therefore,the above problems probably reduce the accuracy of recommendation.In view of the above problems,this paper attempts to propose two improved collaborative filtering algorithms based on user's explicit feedback that is the user's rating and implicit feedback behavior,namely the social tags.The first algorithm is the collaborative filtering based on time,gravitation and preference of item attributes.First,this paper proposed a time decay function based on the time axis of rating on item attributes,which regard it as time weight of user's rating,and construct user-preference of item attributes matrix.Second,this paper also tries to propose a user similarity algorithm based on gravity to measure user's similarity.This paper argues that the user similarity is similar to the universal gravitation between two mass points in physics.The next step is to use the user's rating and user similarity to predict the ratings of target users on unknown and recommend the top-n items.The second algorithm is the collaborative filtering that is based on the Word2 Vec model and social tags.At first,this paper introduces the Word2 Vec language model which is being used to generate semantic social tags word vectors.The next is to,cluster the user's social tags word vectors according to the semantic association.This paper constructs a user-tags category frequency matrix and normalizes it.Moreover,the collaborative filtering will calculate the user similarity and predict unknown scores.Eventually,this paper combines the two improved algorithms together,and meanwhile taking the user's ratings and social tags,which means the explicit and implicit feedback into consideration to generate a hybrid collaborative filtering algorithm to solve the problem of the change of user's interest,sparseness and scalability.It is verified that the proposed algorithm has better performance compared with the traditional collaborative filtering algorithms and other related improved algorithms based on the experiments with the different size of the MovieLens dataset.