Abstract:

There are more knowledge sources than Google, YouTube and Wikipedia around. Thousands of multi-medial contents – knowledge pieces – are appropriate to enrich web-based self-controlled learning on both, fixed (e-learning) and mobile (m-learning) devices. Short breaks can be filled smoothly by consumption of undiscovered video clips, fitting to a personal curriculum covering multiple subjects. Assuming that a big number of free knowledge pieces is already discovered and accessible by the web, there is a strong need for a homogenous classification scheme and personalization to provide flexible guidance by individual recommendation for user self-controlled edutainment.

The full paper will address a new approach in recommendations: stereotypes and collaborative filtering. Next, it focuses to innovative user profile generation and adaptation due to several individual and community factors. A general modular framework, supporting user self-controlled edutainment, including first practical experiences, will be described. Research status and next steps are discussed finally.

While most of the metadata approaches are content centric, poor in reliability and heterogeneous, a innovation project at Deutsche Telekom Laboratories relies on “stereotype-based recommendations”. It offers a scalable mediator between a huge number of knowledge pieces and individual users. Applying the analytic hierarchy process (AHP) algorithm, a set of several dozens abstract stereotypes is generated. A single stereotype is a cluster representative of similar knowledge pieces. Hence, the set of stereotypes is semi-fixed (long term recalculation will be supported). Both, each content and each user are assigned an individual profile. An ordered item list is pre-calculated per stereotype to increase the recommendation performance. A single profile is represented as an affinity vector containing related weights per stereotype. The initial user profile is generated by means of a dynamic questionnaire, presented to the user as a sequence of dependently calculated questions, comparing two proposed knowledge pieces per step. While the user is rating, the weights are adjusted step by step.

Once a profile is created, the user just clicks the “Teach Me” button to get a personal knowledge piece recommendation. The research project combines important advantages, e.g. ease of use; a modular and open system architecture relying on web services and a Hibernate persistence layer; browser based clients for PC and mobile devices without any installation. User profiles are continuously adjusted by the learning community. “Collaborative filtering” is a concept that allows to bundle the experience of several users having similar profiles, while “community-based recommendations” apply the well-known friend-of-a-friend principle. The user is given the chance to actively influence the personal profile by rating (agree/reject) recommended knowledge pieces. The influence of additional factors, implicitly driven by the interaction of community members, like latest access, access frequency or dwell time are currently under investigation.

A field trial within the entertainment domain gave good results regarding the implementation and will open the appliance to various new domains, like education in virtual learning environments (VLE). Dependencies of profile adjustment factors on the framework need to be practically checked. “Sequential recommendation” is an example for future research.

Keywords:

recommendation framework, content based recommendation, collaborative filtering, community.