Recommendation Engine

The Harvest project La Vie has developed a proof-of-concept recommendation system to provide users of VideoLectures.Net with advice on suitable videos for their needs.

One problem created by the success of videolectures is the difficulty that individual users have in identifying the best video for their needs among the vast range of possibilities offered by the site. Each video has a particular mix of content and style of presentation with implicit assumptions about background knowledge and level of expertise of its intended audience. On the other hand the video consumer has an approximate understanding of his/her abilities and material that he/she would like to learn about.

For example, they may have a background in basic classification methods (e.g. SVMs) applied to text, some knowledge of probability theory, but do not know about Bayesian reasoning. He/she would like to learn about Topic Models. The question of which sequence of videos would be most appropriate to help him/her to attain the desired knowledge would also depend on the style of presentation he/she prefers and so on.

The Harvest project La Vie has developed a proof-of-concept recommendation system to provide users with advice on suitable videos for their needs. The system was broken down into six phases. First the text extraction modules uses text mining methods to extract information from the content and meta-data associated with a particular video. The second phase uses information retrieval techniques to retrieve related data from online external resources, i.e. Wikipedia.

Phase three employs automatic speech recognition techniques from the EU transLectures project in order to obtain the automatic transcriptions of videos. The fourth phase involves topic and user modeling from the extracted text in previous phases in order to develop a richer semantic representation. Phase five provides relevant recommendations by linking the enriched semantic representations of users and videos on a basis of a SVM classifier. The final phase collects feedback from users about the operation and effectiveness of the recommendations that have been given. This information is used to update our models.

All these components were integrated into VideoLectures.Net through the user interface that provides visualization and interactivity. In addition, a system update module was developed in order to synchronize our database with VideoLectures.Net, specially for including recently published videos to our recommendation engine.

Currently, the recommender is working on the VideoLectures.Net’s development machine and providing significantly better recommendations, based on a informal human evaluation. It is planned to deploy the recommender into the real VideoLectures.Net web site in the next weeks. This will allow us to perform a real-life evaluation of the La Vie recommender and to assess its quality and benefits against the old one.