Movie genre classification using binary relevance, label powerset, and machine learning classifiers

Multi-label text classification (MLTC) is a technique to categorize texts into more than a single category and used extensively in various real-life problems. Such classifications problems are challenging and dependent on many factors and changes according to the problem. Movie genre classification is a popular multi-label text classification problem as movies may belong to multiple genres at the same time. The major factors used for movie genre classification are based on parameters like movie plot, title, summary, and subtitles. In recent years, some neural networks based approaches are proposed for solving such problems, which turns the solution into resource intensive and time consuming activities. In this paper, we propose a novel method of movie genre classification using a combination of problem transformation techniques, namely binary relevance (BR) and label powerset (LP), text vectorizers and machine learning classifier models. We perform binary relevance task (BR) that converts multi-label classification tasks into independent binary classification tasks whereas label powerset transforms a multi-label problem into a multiclass problem with one multiclass classifier trained on all unique label combinations found in the training data. Further, we apply text vectorizers namely, CV (Count Vectorizer) and TF-IDF (Term Frequency - Inverse Document Frequency) to tokenize the textual data to build a word vocabulary followed by employing various classifiers i.e., Logistic Regression (LR), Multinomial Naive Bayes (MNB), K-Nearest Neighbor (KNN), Support Vector Classifier (SVC) with the combination of different vectorizers and problem transformation methods. To test the effectiveness of these combinations, we use the k-fold cross-validation technique. We construct different combination using problem transformation approaches, text vectorizers and classifier models leading to overall 16 different combinations for classifying movies into appropriate genres. Finally, we evaluate the performance of each combination on publicly available IMDb datasets with target on 27 major parent genres using different performance measures and reveal that the best result is obtained using the combination comprising of label powerset (LP) as Problem transformation approach, TF-IDF as the text vectorizer and support vector classifier (SVC) as the machine learning classifier model with a commendable accuracy of 0.95 and F1-score of 0.86.