Research On Family User-facing Over-The-Top TV Resource Recommendation

With the rapid development of internet resources of the industry and further promote the triple play, fundamental changes are taking place in the traditional TV viewing patterns, new technology is also continue to change the people’s TV screen. The traditional TV service mode has been difficult to meet the diverse needs of home users. Now more and more home users hope to achieve the purpose of entertainment via Internet TV or Internet TV set-top boxes-one plus TV. But it is very difficult to find the resources which Internet TV users is interested in for many users, so it is also very necessary to provide customized Internet TV Resources for home users, making them enjoying the personalized service and improving the user experience.The traditional Internet resource recommendation system mostly recommends for individuals currently, but the whole family is the audience of home TV. Among, most family members is made up of several individuals, as a result, the choosing play collection of the whole family members composes the history viewing of user in the system. Although home users’viewing habits affect each other, the resources of each family member preferences often differ, even completely different, it can not be simply superimposed integration. Therefore, the recommend effect may be counterproductive just according to the traditional recommendation algorithm, not considering the diversity of members’viewing, which directly regarding the home users as an ordinary individual and recommending the traditional algorithm.By analyzing the view history of home users, this article firstly will propose the optimization RFMP model which based on RFM model. It will calculate the value of the contribution of users in the system through the four viewing index of the user. Which are weekly visits, weekly visits of the resources, each resource watching completion rate, as well as measured distribution by hierarchical analysis of the time difference for the last accessing view.It will group the users of the same value behavior as one and conform a first group according to the value, which will provide a basis for the subsequent recommendation model.Meanwhile, in order to identify whether the family is multiple users, provide users with more accurate resource recommendation, we use the algorithm to calculate outlier cycle-resource spacing scoring matrix calculations, based on outliers found in line spacing to identity whether the current user is a member of the home user or multi-individual users. By digging through the home users’viewing habits, combining the three dimensions, watching frequency,watching duration and watching time of the resources for the user, we will obtain the scores for the users of different resources as well as the interest Matrix of each home users, and scored integration based on a single resource belongs to the plate, the user has to each resource sector score range, combined with the resources sector ratings forming periodic interval corresponding user cycle each resource sector-Resource scoring matrix.For different value groups of home users and individual users, individually run association rules of similar interests on each classification dataset and adopt different push policy for each classification. If user belongs to high value of individual users, we get rid of the excessive influence of active users and compute user similarity by cosine correction formula for recommending information by collaborative filtering recommendation algorithm. Otherwise user belongs to low value of individual users, the recommended model not only adopt collaborative filtering recommendation algorithm, but also randomly recommend some not related resources in order to make up for the sparsity of interest matrix and cold boot problems and increasing coverage rate of resources. In addition, if the user belongs to high-value family multi-members, we separate family members to many independent individuals and match watching features user with watching template of every individual, reducing the complexity of recommendation. For individual members fail to match individual users with high value, we will adopt the recommended model; If the user belongs to the low value family multi-members, the interest matrix will be very sparse, using recommendation algorithm directly will reduce accuracy, so this group will be very suitable to adopt comprehensive balanced recommendation strategy by removing the minimum disappointed degree. The advantage of the strategy is that it is not consider interaction of the family members, but also make sure effectiveness of the recommendation. Recommendation based on different value group and home users and individual users is more in line with actual situation. The experimental results show that the model is more stable and effective than the single recommended method. Recommended accuracy and recall rate of the model also has improved.