Research On Collaborative Filtering Based Context-Aware Recommendation Algorithm

As the most widely used personalized recommendation technology,collaborative filtering algorithm makes some success. However traditionalcollaborative filtering algorithm does not consider the time context, treats theinterests of different users in different time equally, ignores that the user’s interestwill change with time and user will forget the items that they visited long time ago,resulting recommendation quality decreased. Currently, collaborative filtering basedon the time-weighted algorithms, mainly uses the ratings time, does not consider thetime the recommendation system recommending items to the user. When the userdoes not produce new behaviors, there is no change in the recommended result list.Ideally, the recommended system should be able to recommend different items to theuser in different times to increase the result list’s time diversity.The main work and contributions presented in this thesis are as follows:(1) Analyze the advantages and disadvantages of the various recommendationalgorithms in detail. It points out that precision and time diversity of the result areimportant to recommendation algorithm.(2) Discuss the dynamic characteristics of recommender systems and theimportance of time context for recommendation.(3) Study the interest variation of different users at different moments and userelative time decay function to decay the item-item similarity, making both theprecision and diversity of the recommended result improved.(4) Study the user memory forgotten theory, and use the simulated Ebbinghausforgetting curve in the process of user’s interest prediction. Then use the absolutetime difference between recommendation time and rating time to weight thecontribution of the item visited by the target user, and use relative time decayfunction to decay the item-item similarity, making the improved collaborativefiltering algorithm can adapt to the changes in user’s interests. This not only canimprove the recommendation performance, but also increase the time diversity ofthe recommended results.