Content-based filtering and personalization using structured metadata

Abstract

Structured descriptions of multimedia content and automatically generated user profiles are used to filter content. Categories OVERVIEWUser profiles that contain compact descriptions of users' interests and personal preferences provide a method for selecting from an increasing amount of multimedia content and for reducing information overload [1,2]. A user profile can be used to (a) filter input content, so that programs or items that the user has shown interest in are presented to the user, and/or (b) request from a content distribution service the programs of interest. Emerging standards such as MPEG-7 and TV-Anytime [3,4] provide an XMLbased, interoperable framework for structured and rich description of multimedia content, along with tools that enable improved user interaction with such content, e.g., tools for representing users' preferences and history of content usage. In this paper, we propose novel algorithms for (i) automatically determining a user's profile from his/her usage history (profiling agent), and (ii) automatically filtering content according to the user's profile (filtering agent). The proposed algorithms are designed to support an MPEG-7 or TVAnytime compliant description framework, although they can also be utilized in a non-standard environment.We present a framework for integrated personalization and management of multimedia content (see Figure 1) which allows non-intrusive observation of user actions over time, enabling collection of consumption-related data without explicit user input. This usage history is subsequently utilized in conjunction with structured metadata that is available for each of the consumed content entities to identify what types of content the user prefers most. Specifically, various features are extracted from the content metadata and usage history database, and are then utilized by the proposed profiling agent that incorporates a fuzzy reasoning system to identify and update the user's preferences. The resulting preference descriptions may include a variety of descriptive attributes, such as genres, actors, channel information, etc., each of which can be assigned a weight reflecting its desirability by the user. Furthermore, the preference descriptions support multilevel, hierarchical grouping of different attributes, and thus provide a highly flexible way of supporting a variety of applications. These descriptions are subsequently compared against the content metadata by the proposed filtering agent, to determine which content entities best match the user's interests. Each individual component of a user preference description corresponds to a test or matching operation against components of descriptions of multimedia content.After individual preference components are tested, the results of these tests are combined into a single composite test result that reflects the user's overall preferences. The filtering agent combines the individual test results according to the hierarchy of the user preferences description, where each parent test is a combination...

On social networks and collaborative recommendation

“…A content-based system selects items based on the correlation between the content of the items (e.g. keywords describing the items, such as album genre, artists, etc., for music tracks) and the users' preferences [6]. However, it is limited to dictionary-bound relations between the keywords used by users and the descriptions of items and therefore does not explore implicit associations between users.…”

An MPEG-7 scheme for semantic content modelling and filtering of digital video

“…Similarly, Martínez et al (2002) describe a system that uses MPEG-7 to catalogue and provide access to multimedia content through annotated variations, in order to permit media content to be offered to different terminals and through different access networks. Ferman et al (2002; propose a profiling agent for automatically determining user profiles from content usage history and a filtering agent for automatically filtering content based on the profiles. Fuzzy inferencing is used to construct and update preferences based on content interactions over a period of time.…”

“…Identifying and retrieving subsets of video content in this way requires user preferences for content to be stated, such that content within a digital video resource may then be filtered against those preferences. While new approaches, such as those based on agents (Wenyin et al 2003), are emerging, filtering in video streams usually uses content-based filtering methods, which analyse the features of the material so that these may then be filtered against the user's content requirements (Ferman et al 2002;Wallace 2002 Content-based filtering is consequently most suitable when a computer can easily analyse the entities and where entity suitability is not subjective (Good et al 1999;Specht and Kahabka 2000). Contentbased filtering is also most effective when the content being analysed has features homogenous with the user's content preferences, since heterogeneous features across the two domains would give rise to incompatibilities, i.e.…”

Metadata-driven multimedia access

“…Discussions of the privacy implications of such personalization systems are included in [7]. Novel algorithms for automatically determining a user's profile from his/her content usage history and for automatically filtering content according to the user's profile are presented in [9]. The profiling and filtering agents proposed in [9] support generation and utilization of MPEG-7 user preferences and usage history descriptions.…”

“…Novel algorithms for automatically determining a user's profile from his/her content usage history and for automatically filtering content according to the user's profile are presented in [9]. The profiling and filtering agents proposed in [9] support generation and utilization of MPEG-7 user preferences and usage history descriptions. The bootstrapping problem of content filtering or recommendation engines (i.e., how to quickly capture a representation of a user's preferences without a complicated dialog or lengthy learning period) is addressed in [10].…”