Content-based recommender systems depend on the viability of metadata in order to filter relevant content to a particular user. Such information can be extracted using automatic and manual indexing techniques. While automatic approaches require the data to be restricted to a specific domain, because they analyze low-level characteristics in order to infer semantic information, manual approaches are expansive and error prone. On the other hand, with the growth of the Web and the possibility of users to create new content and annotations about different items and products, an alternative approach is to obtain metadata created in a collaborative fashion by the same users. However, such information, as keywords, reviews, comments, etc., can have noise, and also be represented by a non padronized and unstructured way. Therefore, this research proposal aims to develop a recommender system based on collaborative descriptions. To reduce the amount of noise and deal with unstructured data, we plan to develop a method which applies different techniques of dimensionality reduction, feature extraction and machine learning in order to obtain a semantically rich and padronized version of data describing those items which will be recommended.
News published in Agência FAPESP Newsletter about the scholarship: