Classifier To Predict The Ratings Of Upcoming Movies

Abstract

In today’s world, movies are released rapidly in all around the world. Therefore, in a particular city, there can be more than one movie in theatres at a time. But people have a busy tight schedules in their life. So, they don’t have enough time to watch each and every movie. Also, due to higher cost of living, people can’t afford to watch all movies in theatres. So, people always try to find the better movies which are worth watching in a movie theatre. Therefore, they try to look for recommendations and non-spoiler reviews from other people. As a solution for this problem, this research project proposes a model. This model was created by analyzing existing movie data, extracting features from it and identifying a relation between the features and the movie rating. There are number of existing research work on movie rating prediction. But all of them have some kind of limitation such as less accuracy, inability to predict rating before movies are released etc. Motive of this research is to overcome most of those limitations. The original dataset was taken from Kaggle and it was updated with some new and missing data retrieved from Facebook, Youtube and OMDb APIs. Most of the existing features were also modified so that they can contribute to the final model in better ways. In this process, the representation of some features were changed, some features were split into multiple features and some features were pruned to have only a selected values in them. Multiple classifier algorithms were evaluated before developing the model and at the end, J48 decision tree algorithm with bagging was selected as it gave the best results. At this point, the output of the prediction model (i.e. the movie ranking) was given as an integer number between 1 to 10. However, in the real world, a person who’s interested in knowing whether a movies is good or not, does not expect such an accuracy. A scale of “Great”, “Ok” and “Poor” would be a good enough measure for this. Taking this fact into consideration, the model was updated to output only 3 values for the rating, namely “High”, “Medium” and “Low” which gave an higher accuracy than earlier.