Towards Be er Recommenda ons for MovieLens Funing Xu Mentor: Max Harper Department of Computer Science and Engineering, University of Minnesota Introduc on Evalua on Results Conclusion MovieLens is a website that helps users find be er We evaluated around 4000 users with at least 40 ra ngs who logged in In this study, results showed when we blend movies built by Grouplens Research Lab. It is mainly within six months. Under different blending strategies both in age a li le popularity and age, the number of using Item-Item collabora ve filtering algorithm to condi on and popularity condi on, the top 20 movies are evaluated on movies in wish list will increase and RMSE will generate predic on score for movies. Currently, the selected metrics. decrease. The NDCG metric showed that orders of recommenda ons are only based on movies’ blending of popularity and age only have a predic on score. The wish list metric measures how many movies in user’s top 20 minor nega ve impact on recommenda on recommenda ons are also in this user’s wish list. The Root mean square quali es. In this study, we tried to blend movie age and movie devia on (RMSE) metric is a typical recommender system metric which popularity to the movie predic on score with different measures the accuracy of recommenda ons. The Normalized discounted The results indicates when we blend strategies. Furthermore, we evaluated different cumula ve gain (NDCG) measures the quality of recommenda ons popularity and age to original predic on blending strategies with metrics in the filed of varying from 0.0 to 1.0, with 1.0 represen ng the idea value. The score, the recommenda ons become less recommender system and informa on retrieval. following charts show the results of our study. personalized. Meanwhile, more recommended movies appear on user’s wish list, which shows that recogni on of recommenda ons is promoted.
Selected Reference
1. McNee, Sean M., John Riedl, and Joseph A. Konstan. "Being accurate is not enough: how accuracy metrics have hurt Figure 1: wish list metric with popularity Figure 2: wish list metric with age recommender systems." CHI'06 extended abstracts on Human factors in compu ng systems. ACM, 2006. Methods 2. Amatriain, Xavier, and Jus n Basilico. "Ne lix recommenda ons: beyond the 5 Two a ributes of movies are chosen as blending factor, stars." Ne lix Tech Blog 6 (2012). which are last year popularity and movie release date.
Popularity of a movie is the number of ra ngs it received.
Since the value of predic on score, popularity and movie Figure 3: RMSE metric with popularity Figure 4: RMSE metric with age release date has different scale, we first convert these value according to percen le rank. Therefore, all of these values are under the same scale of zero to one. Acknowledgement
In the end, we have two groups of new score which This work was founded by the University of decide the order of recommenda ons as follows. Minnesota Undergraduate Research Opportunity Program(UROP). new!score1! = !A! × !prediction!percentile! + !B! × !popularity!percentile new!score2! = !A! × !prediction!percentile! + !B! × !movie!release!date!percentile A! + !B! = !1 Figure 5: NDCG metric with popularity Figure 6: NDCG metric with age