Research and review of movie recommendation Websites
In recent years, tags have been widely used in various recommended Websites, making them an effective recommendation tool. Tags are attached to digital objects to describe the keywords of the object, but they are not part of the formal classification system. It is a freely chosen keyword, but a simple and powerful tool for organizing, searching, and exploring object resources. It has many advantages: first, it directly reflects users' feelings about a movie. Unlike user reviews, the label has no strict organizational structure, but contains rich and clear theme information, which avoids the cumbersome of user comments. Similarly, the label can more flexibly store the user's feelings than the movie type information; in addition, the label behavior forms a three dimensional switch based on the user, object and label. In addition to traditional useritem relations, itemtag and usertag correspond to each other. Because of these advantages, tags are becoming more and more popular and widely used in personalized recommendation. This paper focuses on recommending technology in film recommendation, mainly introduces and analyzes Jinni, IMDB, Criticker three mainstream movie recommendation sites, and summarizes the basic information of the three major Websites, and compares and analyses the mechanism of the film retrieval and recommendation and the features of the Websites in detail. Furthermore, in view of the cold start problem in collaborative filtering recommendation, a simple experiment is conducted on two Websites with strong recommendation mechanism and the results are analyzed.
Jinni can search according to a series of parameters such as the plot, duration, location, awards, keywords and so on. The search tool used by Jinni is called Movie Genome. The popularity of Movie Genome is very high. Google TV also uses this tool to achieve personalized retrieval. The Movie Genome of Jinni contains 2200 different parameters, including type, background and atmosphere. It can find the exact parameters to describe a movie based on the user's comments and other information. Jinni classifies movies by using a large number of parameters contained in Movie Genome and explores user preferences and the degree of correlation among users. First, users should register an account and generate personal folders to collect user behavior and interest information. Users can score at least 10 movies and Jinni can make recommendations. Users and film groups Jinni divide users into 12 categories: socio political, intelligent, heroic, heroic, individualist, strategist, realist, idealist, extremist, imaginative artist, drama enthusiast, sitcom enthusiast and suspense enthusiast. By categorization of users, Jinni can obtain user preferences for each type of user interest and behavior to further clarify the user's preference type. And the user's rating of the movie is divided into 10 levels, from low to high to be terrifying, bad, boring, disappointing, general, good, very good, amazing and absolutely unmissed. Avoiding the traditional scores, and using the emotional score can clearly show the user's interest in a movie, and eliminate the problem of the vagueness of the interest of the conservative scoring users in the traditional score. For example, a conservative score of 6 users has already been satisfied with the movie, which is equivalent to 8 or more of the average user. But using emotional comments, each level represents a clear degree of preference, thus solving the previous problems. The preference location recommendation system analyzes user's data according to user's score and behavior information, and gets the user's preference, so as to recommend. According to a user's rating of a movie, Jinni can see whether the user likes the movie, and then draft the user preferences. Jinni divides the crowd into 12 major categories. Based on the score of users in each of the major categories, Jinni uses six different entries to describe user preferences. For example, strong leaning indicates that users are strongly inclined to this type, while not at all indicates that the user is not the type at all. For the formulation of user preferences, users need to score at least 20 movies under the same category, otherwise it is difficult to give results. Because it is very important to determine the degree of interest for the user in a classification, the error evaluation will directly remove the user out of the type, resulting in a great deviation in the later recommendation. The more the existing film scores, the more complete the ability to test the user's interest in this type of film, thus improving the system's judgment of the user's interest and avoiding the preference error. Finally, based on the comparison with other users, users can understand their relevance with other users and choose whether to follow them. If the user chooses to follow, the movie that the follower has recently watched will have an impact on user recommendation. Jinni will display closely related other users' recent movies and most people's ratings of the movie in user folders. When the user enters the Jinni recommendation page, the user's recommendation list is generated according to the user preferences; in addition, according to the recent activity information of the similar user group, a higher grade film is selected to produce another recommendation list.
IMDB uses multiple searches, which use different keywords to filter the demand information. IMDB can search titles, abstracts, types, keywords, movie companies, plots, etc., and also has advanced search. Advanced search includes advanced title search, advanced name search, and collaborative overlapping search. Advanced title search can take multiple factors such as movie type, age, voting number, score and so on into consideration. Advanced name search can search according to the actor's sex, date of birth, name and height. Cooperative overlapping search can achieve simultaneous retrieval of two titles or two names. Enter "Apocalypse Now (Modern Apocalypse)" and "The Godfather" in the Title Synergetic search, and search for people who appear at the same time in two movies, actor Robert Duvall and director Francis Ford Coppola. In the name synergetic search for the simultaneous input of director Frank Darabont and actor TimRobbins, there will be a search for a film of two people, such as the shawshank redemption.
Criticker's movie scoring method is very distinctive, it uses Taste Compatibility Index (TCI) to locate user preferences. When users establish their accounts, they can score movies for the Website. According to the user's movie rating, Criticker will automatically set the rating for the rating movies, ranging from 1 to 10. The rating of the system is to rank all the movie ratings of the user, and then to give the grade scores from high to low. Therefore, even if two users give the same score to the same movie, the grading is not necessarily equal. For example, A users scored 65, 75, 85, 90, 95 for 1~5 five films, while B users scored 65, 55, 45, 35, 25 in the same five films. The movie rating of A users was 1, 2, 3, 4 and 5 respectively, while those given B were 5, 4, 3, 2 and 1. Therefore, A and B users scored the same score on the first movie, but the system's grade scores were quite different. In this way, each user will produce a class library of his own movie, and any two users will have a difference in the difference in the grade score of the same movie. When the average of all the differences can get two users' TCI: TCI index less than 3, the interest of the two users is similar, so the movie is recommended. When TCI is greater than 3, the two users can think of no common interest, so the choice and preference of both sides will not affect each other.
At present, the collaborative filtering recommendation system is not fully mature. From the movie recommendation Website alone, the problem is as follows: many movie recommendation sites still stay in the information retrieval level, and do not do the real personalized recommendation, and do not solve the personalized problem of the user specific needs; for personalized recommendation, most films push the movie. There are no effective methods to solve the main problems such as cold start and data sparsity, or the accuracy of the method is not high. In the case of film information redundancy, the accuracy of extracting effective information in collaborative filtering is not high. In order to solve the cold start problem better, we should consider the following aspects: first, we should solve the cold start problem of the new film, that is, how to realize the accurate recommendation of the new film. The recommendation system needs to increase the user's information more and faster than the original information of the film, which requires a corresponding increase in the user's opportunity for the evaluation of the new film. Secondly, for the new user problem, the basic attributes such as user interest orientation can be added to the user registration process, so that the system can obtain interest information immediately when the new user is registered, so that a simple recommendation can be realized on the basis of interest matching when the user is cold starting. To achieve this recommendation, we need to construct a "user resource" model to connect users and interest resources. For new users, they can be assigned to the corresponding user model according to some important attribute features, and the new film is also assigned to the corresponding movie category model by using the Bias classification method. In addition, the unique advantages of the label make it a development direction of the recommendation system. It can effectively use the simple and powerful information of the label transfer. It can not only calculate the similarity of the users, but also calculate the similarity of the film, thus establishing the data model of the user - label - film three, and realizing the final recommendation.
In order to improve the accuracy of the future film recommendation system and to solve the main problems, it is necessary to strengthen the use of effective collaborative filtering, and pay more attention to the problems of cold start and data sparsity in personalized recommendation, and in the calculation of user similarity and movie resemblance, it is necessary to choose more information. Be careful. Because of its short and clear definition and strong relevance between users and movies, tags will probably achieve a major breakthrough in film recommendation.
