| Summary: | Typically, when a human being watches a video, different sensations and mind states can be stimulated. Among these, the sensation of fear can be triggered by watching segments of movies containing themes such as violence, horror and suspense. Both the audio and visual stimuli may contribute to induce fear onto the viewer. This dissertation studies the use of machine learning for forecasting the emotional effects triggered by video, more precisely, the automatic identification of fear inducing video segments. Using the LIRIS-ACCEDE dataset, several experiments have been performed in order to identify feature sets that are most relevant to the problem and to assess the performance of different machine learning classifiers. Both classical and deep learning techniques have been implemented and evaluated, using the Scikit-learn and TensorFlow machine learning libraries. Two different approaches for training and testing have been followed: film-level dataset splitting, where different films were used for training and testing; and sample-level dataset splitting, which allowed that different samples coming from the same films were used for training and testing. The prediction of movie segments that trigger fear sensations achieved a F1-score of 18.5% in the first approach, a value suggesting that the dataset does not adequately represent the universe of movies. The second approach achieved a F1-score of about 84.0%, a substantially higher value that shows promising outcomes when performing the proposed task.
|
|---|