Learning and Explanation of Extreme Multi-label Deep Classification Models for Media Content

Abstract

Providing rich and accurate metadata for indexing media content is a crucial problem for all the companies offering streaming entertainment services. These metadata are typically used to improve the result of search engines and to feed recommendation algorithms in order to yield recommendation lists matching user interests. In particular, the problem of labeling multimedia content with informative tags (able to accurately describe the topics associated with such content) is a relevant issue. Indeed, the labeling procedure is time-consuming and susceptible to errors process as it is usually performed by domain experts in a fully manual fashion. Recently, the adoption of Machine Learning based techniques to tackle this problem has been investigated but the lack of clean and labeled training data leads to the yield of weak predictive models. To address all these issues, in this work we define a Deep Learning based framework for semi-automatic multi-label classification integrating model prediction explanation tools. In particular, Model Explanation techniques allow for supporting the operator to perform labeling of the contents. A preliminary experimentation conducted on a real dataset demonstrates the quality of the proposed solution.