Research Of Identification Smart TV’s Family Member Based On LDA

With the rapid development of Internet of Things, the concept of smart home and its product is popular. Smart TV as one of the most important consumption products in our life, recently get rapid development and popularization. One key advantage of smart TV over the traditional terrestrial TV, satellite TV and cable TV is that smart TV as a social media service via TV and social networks through which TV users exchange their experiences about programs that they are viewing. Smart TV over the traditional terrestrial TV is that it makes the TV viewing experience more interactive and personalized. For smart TV service, two technical aspects are envisioned:grouping of similar TV users to create social TV communities and recommending TV programs based on group and personal interests for personalizing TV. Many efforts have been made for personalizing smart TV services. A key challenge in analyzing smart TV user behaviors is that smart TV do not like mobile phones, computers or other electronic products which generally only used by one user, smart TV is used by a family, the members of a family are not sure, may be composed of a single, also may be generations of people live together. And the time of each one watching programs in the family are not sure. Thus infer the member of the family and the main active period of each member in the family is the basis of to provide personalized service for smart TV consumers.Based on the above demands, in this paper we propose a method to identify family members of smart TV based on LDA model. LDA is a generative probabilistic model for collection of discrete data such as text corpora, it has three-level hierarchical Bayesian structure (document-topic-word), we use the LDA model to analyze the text description of media.First, we select medias from video database and process the text information of media like title, tags, summary by using Chinese Word Segmentation, retain the relevant part of words(noun, verb, adjective, adverb) as the LDA index file, the LDA model base on the input information for training, generate various topics, the key words for each topic and probability distribution of each film on the topics.Then we select family history record in weekdays and divide time into 24 periods. According to the history of smart TV family watching and combine with the probability distribution of each film on the topics obtain the probability distribution of each user on the topics in each period; According to the interested topic vector on family in each period of time, through the method of experts’definition to determine the mapping relationship between interest topic vector and user attribute and establish a mapping table; Each family’s interest topic vector compare with the table of interest topic vector-user attributes to infer the members of family, the active time of each user and the information such as the dominant member in every family.Finally we implement the method of identifying family members about smart TV in the real environment. And we propose a method to improve the identification accuracy by setting threshold condition to get rid of noisy data. The experimental result shows that this method is an effective way to improve the recognition accuracy.This paper also propose another method to analyze smart TV family member. Using multi-class support vector machine to analyze smart TV family member. The user characteristic vector as multi-class support vector machine(MSVM) training set, get a classification model, apply the classification model to predict test set and get accuracy. This method respectively compare two kinds of classification function effect and the judgment accuracy of sorting machine and kernel function combination.