zMovie CSE 401: Senior Design Jing Chen, [email protected] Faculty Advisor: Sudipto Guha Anant Jatia, [email protected] -1- zMovie: The Movie Recommendation Engine I. ABSTRACT zMovie adds a whole new dimension to the movie watching experience by providing real-time personalized movie recommendations to users. It takes a collaborative social-networking approach where a user’s own tastes are mixed with that of the entire community to generate meaningful results. Most existing movie services like IMDB (www.imdb.com) do not personalize their recommendations but simply provide an overall rating for a movie. This significantly decreases the value of each recommendation as it does not cater to the individual movie preferences of the user. Unlike these systems, zMovie’s Recommendation Engine will continually analyze individual user’s movie preferences and recommend custom movie recommendations. The overall goal is to ease the movie discovery process. zMovie is purely a movie recommendation service in that it offers a list of movie suggestions based on previous user ratings. zMovie is designed not to search for movies but to discover them through our recommendation process. zMovie will allow users to rate movies they have seen. This data is then analyzed, and recommendations are then returned to the user. The core of our project, zMovie’s recommendation algorithm, is based on a cluster-smoothed collaborative filtering algorithm [2]. We have refined and tuned the parameters around this algorithm by comparing our predicted ratings against actual ratings using in-sample and out-of-sample techniques as well as analyzing live user feedback.zMovie CSE 401: Senior Design Jing Chen, [email protected] Faculty Advisor: Sudipto Guha Anant Jatia, [email protected] -2- II. RELATED WORK EXISTING PRODUCTS AND SYSTEMS Current Social Networking World Internet social networking sites, which began in 1995 with Classmates.com, have surged in popularity and use through word-of-mouth advertising. Since then, a wide range of virtual communities have formed serving different purposes and targeting varying niche audiences: ProfessionalsActiveRain (real estate)Ecademy (business)LinkedIn (business)Orkut (Google)Cultural CommunitiesBlackPlanet.com (African Americans)Cyworld (South Korean)Hyves (Dutch)IRC-Galleria (Finland)iWiW (Hungary)LunarStorm (Sweden)MiGente.com (Latinos)AcademicBebo (schools) Classmates.com (public)Facebook (public)Xuqa (colleges)BlogsBlurtyLivejournalXanga VoxWindows Live SpacesTabulasPhotosFlickrMyPhotoBucketWebshotsPicasaRelationshipsMatch.comMultiplyMusicLast.fmMystrandsBoltGoldmicMusic ForteMymidishareReviewsTribeRateitallChowhoundYelpVideoYouTubeBlinkxAkimboDave.tvBrightcoveSocial Networking WorldRelationshipsBookmarkingdel.icio.usdiggredditMoviesYahoo! MoviesMovieLensFlixsterBlockbuster/Netflix Social Movie Platforms In particular, we’ve chosen to explore the movie niche as this is an area where our project can provide significant improvements compared to existing products and systems. Traditional movie websites (IMDB, AOL Movies) function by proving global user ratings on movies in their database. Movies are categorized by metadata such as genre, era, directors, and so on. Users can search for movies, browse lists and read reviews written by critics or other users. However, most of these services lack any personal recommendation system and haven’t taken advantage of social-networking communities or crowd wisdom. Some websites, such as Blockbuster, do provide individualized recommendations based on a user’s ratings but do not include any social networking component. Yahoo! Movies goes further and uses personal ratings to suggest movies currently playing in theatre, on TV, and out on DVD. It also drawszMovie CSE 401: Senior Design Jing Chen, [email protected] Faculty Advisor: Sudipto Guha Anant Jatia, [email protected] -3- upon its vast user base to give lists of similar movie fans, their ratings, and reviews. Other movie sites, like Flixster, take a different approach. Flixster forms web-based communities around movies and suggests movies to watch based on what your friends have rated. Recommender Systems Two conventional paradigms applied to recommender systems and user preference predictions are collaborative (CF) filtering and content-based (CB) filtering [1]. Collaborative filtering makes recommendations for a given user based on aggregating rating information of similar users in a historical database. On the other hand, content-based systems provide recommendations by comparing representations of the content contained in a given item (i.e. book, movie, song) with representations of content that match a given user profile. While CB systems can characterize users more uniquely, CF systems have been more successful because they not only do not require content to be associated with items but also can provide recommendations that are relevant to a user without having them contain content from the user’s profile [1, 2]. For these reasons, we have chosen to focus primarily on CF systems. There have been two primary kinds of CF systems (See Appendix A for illustrative diagrams of each recommender system): 1. User-Based CF: In User-Based CF [13], a target user’s choices are compared with other users in the database to identify a group of “similar minded” people. Once this group is identified, highly rated content from the group are then recommended to the target user. Limitations of this include:  Bias towards what has already been recommended or chosen - frequent recommendation of most popular items and poor new item discovery tool  “Cold start” problem - items must be chosen by users before recommendation can be made; inability to recommend new content  Potential poor quality in recommendation - system only accounts for user’s pattern of choices (ratings on items) with no understanding of underlying content behind data (attributes of an item)  Extremely data-intensive - requires large number of user choices and ratings, often 30 or more, before reasonable recommendations can be made  Difficulty to analyze in real time - especially as number of users in the database grows 2. Item-Based CF: Item-based CF [13] examines each item on the target user’s list of chosen/rated items and finds other items in the choice set that seems similar to the item. In this case,