한국통신학회 2010 년도 추계종합학술발표회 NavProfiler for Multimedia Personalization in IPTV Nalin Chakoo O , Saewoong Bahk School of Electrical Engineering and Computer Science, Institute of New Media and Communication, Seoul National University {nalinchakooo, sbahk}@netlab.snu.ac.kr Abstract User Profile forms the basis of personalized content recommendation in Recommender Systems. These systems value user- profile as a true representative for the user's interests. However, current methods used to construct user-profile on user’s implicit behavior are limited. In this paper, we propose a mechanism to capture and evaluate user ’s navigational and content browsing pattern such as on IPTV content, and integrate extracted information into user-profile. The method enriches user- profile vector for accurate similarity evaluation among IPTV peers. We explain the design of profile agent, NavProfiler, which captures meta-data from user’s selected content, accurate viewing time and cache references. We extract the interests automatically without prior training. Our approach expands the vector components of the user-profile for better comparison and quality recommendations. Ⅰ . Introduction Consumer Electronic (CE) devices are becoming powerful with the stupendous growth in hardware and software technologies. Compounding to these advancements, expanding infrastructure makes network-capable entertainment devices very popular. Users are creating, storing, and sharing more and more digital content through these internet capable devices. Search engines lessen the task of searching content by filtering contents that match user’s query but again in an inadequate capacity. Additionally, most of the search engines fail to take into account user’s taste. To fine tune user experience, the concept of recommender systems was introduced. A recommender system can be expressed as a type of information filtering (IF) engine [1] that recommends contents. The system creates a user-profile based on the content viewed by the user, compares the user-profile with some reference characteristics and finally computes the user’s likeness for a particular content. These characteristics may be extracted from the meta-data associated with the content i.e. Content Based technique or meta-data from other users i.e. Collaborative Filtering. The capability of a system to generate recommendation to a particular user means that it must be able to deduce what that user requires and be capable of aggregating similar users as accurately as possible so as to get finer result sets. This aggregation process is based on similarity evaluation which further is computed using user-profiles. Ⅱ . Related Work Recommender system in consumer electronic devices, such as IPTV, suggests contents based on the opinions of other users or in other words, user-profiles [1]. These recommendation services on IPTVs can be fitted in a user behavioral model where the system captures accurate user likeness factor of multimedia content items. The likeness factor can be quantified for various contents by associating interest weight according to the navigational and browsing pattern of the user. The associate interest weight of an item is typically expressed as a numerical value reflecting the user’s level of interest in that item and can be integrated into the user profile which is generally a vector of keywords and weights. In order to calculate similarity, it is a common practice to represent this user profile in the bag-of-words (BOW) format [2] which is a set of weighted terms that best describe the entity so that the similarity between two entities can be computed using some standard approaches such as Cosine, Pearson, Jaccard Coefficient [3] etc. In this paper, we describe a novel method that captures user’s navigational pattern in IPTV viewing into user-profile. This approach to enhance user-profile has not been used in previous work. Therefore, the resulting user-profile, an embedded tag extractor, facilitates better personalized content recommendation. Ⅲ . System Design Formally, consider a user-profile can be viewed as a mapping of users and multimedia tags to a set of interest weights, we suppose a set of users U = {u 1 , u 2 , · · · , u m }, and a set of multimedia content tags, MCT = {mct 1 , mct 2 , · · · , mct n }. The profile for a user uU is therefore an n- dimensional vector of ordered pairs: u (n) = {(mct 1 , iw u (i 1 )), (mct 2 , iw u (i 2 )), ··· , (mct n , iw u (i n ))}, where mctMCT and iw u is a interest weight (iw) function for user u, assigning weights to multimedia tags in MT. Updating user-profile typically involves two stages: a) User’s content watching context: b) Tag search and collection Our approach is designing a context-aware navigation agent in the IPTV. The agent performs two steps: First, the navigational agent identifies channel navigation and retrieves tags from the closed captions (CC) and EPG source as shown in Fig 1. Figure 1: User Profile Enhancement Process Second, NavProfiler, developed over IPTV Architecture