Deep learning music information retrieval for content-based movie recommendation

Summary

A recommendation system is a system that provides suggestions to users for certain resources like movies. Movie recommendation systems usually predict what movies a user would like based on the attributes present in previously liked movies. Content-based movie recommendation systems are important because they base their recommendations mostly on the content of the movies such as music, speech and subtitles. Music, in particular, is an import aspect of a movie, as, in many cases, it may reach the audience emotionally beyond the ability of picture and sound. In this thesis, both machine learning and deep learning are used to extract valuable information from movie music, in order to create a good base for a content- based recommendation system. Towards the end, a dataset is built containing the predicted values of the aforementioned information of 145 movies of 12 different directors and a collection of metadata. At a first stage, a simple music detector based on SVM classifier and hand crafted features is trained and applied to movies in order to detect music segments of each movie. Then 4 deep learning models are trained to classify 4 music at- tributes, namely: danceability, valence, energy and musical genres. In addition, an interactive web application was created with dash Python web framework and plotly for data investigation and a series of plots were produced through a mining procedure in the aforementioned dataset.