Pre-release forecasting of imdb movie ratings using multi-view data

Internet users are no longer the ones who only follow the content, but at the same time they have become the ones who create the content itself. Especially, the social media has become a crucial part of everyday life. This results in having a huge digital footprint for the people on the internet either willingly or unintentionally. People provide enormous data such as thoughts on the current issues, comments on an owned product, opinions about a recently released movie, experiences about the visited places, the political views, and so much more different topics thanks to sharing almost everything on these social media platforms. The existence of such large-scale data attracts many researchers from different areas in order to produce useful outcomes by analyzing it. If such an analysis system is designed properly, it could produce more accurate and precise results than the traditional methods like analyzing the output of the surveys using statistical techniques. These systems could be used in everyday fields such as predicting the urban traffic, detecting the news popularity, the likability of a new product coming onto the market, as well as in the fields like financial sector and politics. Making inferences using such systems have become quite important also for the ever-growing film industry. The film industry has a very important share in the global market, with billions of dollars revenue per year. Therefore, it is very important to increase the profit by predicting the movie success before its release. However, most of the total revenue is generated by the only a very low percentage of films. This also means that most films make a loss. In such a billion-dollar industry, it becomes very significant to make a profitable film. From the film industry's point of view, achieving the knowledge of which factors make a film high grossing is very crucial. Detecting those factors before the release of the movie could help the film producers and the other decision-making people to make more appropriate decisions such as deciding the theater's count in which the movie plays or increasing the promotions to raise the revenue. For instance, it could be expected that the revenue will increase if more theater rents for a movie that is predicted to become successful since more audience could watch this movie. On the contrary, deciding to rent less theater for a movie predicted to become unsuccessful could help to reduce the loss. For all these reasons, predicting the box office gross, hence the success of a movie is a very important step in the film industry. It could be assumed that there is a correlation between the box office gross and movie success. As the success of the film increases, the box office gross also increases. However, this prediction is very challenging because it is difficult to distinguish the factors making a movie successful, and detecting the people's opinion is a very subjective issue. Such predictions could be made by using traditional methods like the Hollywood Stock Exchange. However, these conventional methods are incapable to carry out such a complicated task. Recent studies showed that data mining and machine learning methods yield more accurate results. Thus, forecasting the success of a pre-release movie has become an intriguing research area also for the computer scientists. Machine learning is a computer science field to improve performance on a specific task using existing data, which is also called training data. The aim is to create mathematical models to make inferences from these past experiences. Machine learning heavily makes use of the methods from statistics in order to build these models. In other words, machine learning involves the optimization processes of the parameters of these mathematical models using the training data. Most of the time machine learning is associated with data mining. Data mining is the application of machine learning methods on large datasets. There is a wide range of application fields of data mining, from detecting the frauds in the finance sector to predicting the house prices for real estate agents; from the medical researches to autonomous vehicles. Predicting movie success for the film industry is one of these various fields too. Predicting the audience ratings of movies before their releases underlies this study. It could be also said that the aim of this study is to build an effective predictive model to solve this regression problem. It is a regression problem because the predicted movie audience rating value is a scalar value rather than a discrete value as in classification problems. Although there are different studies in this field, some of them only utilized from metadata of the movie, whereas recent studies focused on gathering these metadata with social media data to improve prediction results, it is still needed to make extensive analysis and increase performance for predictions. Also, many of the studies explored the box office prediction performance either treated it as a classification or regression problem. However, there were few studies on predicting the taste of movie audiences. Furthermore, many existing studies did not collect enough and recent data. This study aims to take previous studies a step further by finding a proper way to bring together different data sources and increasing the performance results by utilizing different machine learning methods such as Factorization Machines. In this study, it is aimed to explore the use of Factorization Machines in order to predict movie success by predicting IMDb ratings of newly released movies and compare it with different machine learning methods by combining movie metadata with social media data. Also, it is investigated which features have more impact on the movie rating prediction process and how these features coming from different data sources could be used together. Several prediction algorithms were used to solve the movie audience rating prediction problem and to compare the results with the proposed Factorization Machines approach. Linear Regression, Support Vector Regression, K-nearest Neighbors Regression, Multilayer Perceptron Regression, Random Forest Regression, and Gradient Boosting Tree Regression are the predictions algorithms that were used for comparison throughout this study. In this study, a novel approach was developed based on Factorization Machines which combines movie metadata with social media data in order to improve the prediction accuracy of movie ratings before its release. Factorization Machines are an extended version of factorization models. Factorization Machines could combine the advantages of typical machine learning approaches like Support Vector Machines and Linear Regression with factorization models in order to get rid of the drawbacks from these methods. Factorization Machines could be applied both classification and regression problems. There is not an existing dataset which includes the movie metadata and social media data as required. Also, there is not any framework to gather movie data from different sources. Therefore, a framework was developed to gather data from different sources and generate a custom database as the first step of this study. It becomes possible to gather more data by adjusting the input parameters like release date range, language, and region thanks to this framework. Many different data sources such as TMDb, IMDb, Twitter, and Wikipedia were used. These sources provide APIs in order to access their data with specific parameters instead of providing the data with easy-to-use certain text file formats. Therefore, different wrappers were developed for accessing these APIs throughout this study. In this study, it was also explored that how using different feature sets together affects the prediction performance, while exploring the effect of using Factorization Machines approach on this regression problem. Therefore, the features were grouped by their data sources, metadata and social media, and their data types, continuous and discrete values, in order to decide which group of data results in better performance. After all, 6 different feature sets were used to feed the models according to this grouping process. Multivariate Linear Regression is used in many previous studies since it is easy to implement and interpret. Therefore, Multivariate Linear Regression method is selected as the baseline method for performance comparisons with Factorization Machines. In order to calculate performance metrics, 10-fold cross-validation was used by randomly dividing the training data. The performance metrics were calculated as the average of the metrics from these folds. The mean square error (MSE) and coefficient of determination (R2 score), both are the most common performance evaluation metrics, were used as the performance metrics. The best performance metric was achieved an MSE 1.19 and R2 score 0.88 by using the Factorization Machines approach with both continuous and discrete features together for movie metadata and social media. Although there is an improvement by using Factorization Machines approach compared with Linear Regression approach, it is still needed to be enhanced. The experimental results show that Factorization Machines approach increases the performance using continuous and discrete metadata features together. The combination of movie metadata and social media data slightly changes the performance for predicting movie rating outcomes. Factorization Machines approach has also promising results for movie rating prediction as it can characterize latent factors for movie metadata and social media data. Our proposed method improves the prediction accuracy compared to the traditional methods.