Font Size: a A A

Collaborative Filtering In Personalized Recommended Systems

Posted on:2018-06-28Degree:MasterType:Thesis
Country:ChinaCandidate:A Q ZhouFull Text:PDF
GTID:2417330596990104Subject:Applied Statistics
Abstract/Summary:PDF Full Text Request
Personalized recommended systems are very important to solve the Information Overload problem,where the most common method is collaborative filtering algorithm.Collaborative filtering helps users to discover items which they may like through analyzing their behaviors in the history.The basic idea of collaborative filtering is recommending users items liked by other similar users or items similar to other items liked by the same user.This paper investigates collaborative filtering through improving similarity measures,evaluation criteria and filtering methods and conducts an empirical experiment to verify these improvements.First,we propose distance-based similarity measures to address some shortages of traditional similarities in measuring the absolute differences between ratings.The empirical results indicate distance-based similarities may outperform traditional ones in some aspects and possess different properties.Second,to get more accurate reflection of need in reality,we propose three new evaluation criteria: Precision,Recall and F1,which reflect the accuracy of recommendation.The results suggest the new criteria are different with traditional ones to some extent.Then we consider significance-weighted similarities based of the numbers of common ratings to strengthen the reliability of similarities.Generally,performances are improved but we also observed some exceptions.Finally,we propose a high-similarity filtering method to consider objects with high similarities in priority and get improved results relative to the KNN method.
Keywords/Search Tags:Recommended System, Collaborative Filtering, Similarity Measures, K-Nearest Neighborhood
PDF Full Text Request
Related items