1- The audiovisual features from the trailers had to be extracted manually to be used as inputs. The choice of audiovisual features is such that they are related to certain metadata features
2- The author investigate the usefulness of the selected audiovisual features in explaining certain metadata features. Linear SVM was used to predict the category of genre and MPAA rating using audiovisuals. Then for continuous metadata features (budget, movie runtime, etc), linear regression was used. It was found that some metadata can be replaced with audiovisuals. For example, movie genre, it can be easily predicted from the movie trailer.
3- After investigating the usefulness of audiovisuals, the author proceeded to find the correlation between the extracted audiovisuals and the opening weekend gross of the movie. Results from this are in "Results of Study"