Motion picture industry has constantly growth in last decade. Still, this is a tricky business and box office failures could happen. Because of this, predicting the success of a movie, a task known as box office forecasting, is a problem worth studying. Forecast the success or flop of movies is a hard task, which involves the study of several variables, like cast, genre, MPAA, social network marketing, interest by the audience and so on. In many cases, such data is hard to collect, and, in many countries, there is a lack of data sources on movies. In this thesis we propose a new idea to solve the problem of scarcity of data in the field of Box Office forecasting. Traditional models use machine learning and statistical analysis to predict the gross of movies. Such studies usually analyze the US market as it is easier to collect data in a big market like America. In this study we propose to use transfer learning. The idea behind transfer learning is the following: use the knowledge from a source domain, which could be the US market, and transfer that knowledge to a target domain. Because of this, the research question that drives this study is the following: Can transfer learning improve the quality of predictions in a situation where few data is available? We focused our study on the Swedish market, which is an example of country where information is hard to find. We collected a data set of Wikipedia page views from the Swedish market and, using machine learning and statistical analysis, we developed a model able to forecast the box office premiere gross of a given movie. We also collected the Wikipedia page views from the US market and applied transfer learning methodologies in order to improve the performance of Swedish models. In transfer learning terms, this means that the US is the source domain and Sweden the target domain. We finally compared the Swedish-only and Swedish-US models to see if transfer learning is beneficial or not. Results show that the best performing model trained only on Swedish data is able to achieve a R2=0.83. The best performing model trained with transfer learning and using both US and Swedish data achieve an R2=0.82. The conclusion of this study is that transfer learning did not improve the performance in our context. These results are only related to this study. We believe that transfer learning was not beneficial to our study as the two studied markets, Sweden and US, behaves differently when it comes to movie selection.
Nell’ ultimo decennio l’ industria del cinema è costantemente cresciuta, anche se flop cinematografici possono comunque avvenire. Per questo motivo sarebbe utile riuscire a predire il successo di un cinema. Prevedere il successo o fallimento è difficile e richiedo lo studio di molte variabili come il cast, il genere, i divieti, il marketing e l’ interesse dell’ audience di riferimento solo per citarne alcuni. In molti casi è difficile raccogliere queste informazioni e, in molti paesi, queste informazioni non sono nemmeno disponibili. Questa tesi propone un nuovo approccio per sopperire al problema della scarsità di dati disponibili nel dominio dei cinema. I modelli tradizionali usano tecniche di machine learning e analisi statistica per prevedere l’incasso al botteghino dei cinema. Questi studi utilizzano solitamente dati provenienti dal mercato americano poiché’ un grande mercato come l’ America offre una quantità maggiore di dati. In questa tesi proponiamo di usare tecniche di Transfer Learning per sopperire alla scarsità di dati. L’ idea alla base del transfer Learning è la seguente: usare la conoscenza disponibile in un dominio sorgente ( nel nostro case i dati raccolti dal mercato US) e trasferire la conoscenza implicita nei dati in dominio target. In questa tesina il mercato svedese è il dominio target in quanto esempio di paese dove è difficile raccogliere informazioni sui cinema. Abbiamo collezionato un data set che raccoglie, giorno per giorno, il numero di visite alla pagina Wikipedia di un film e poi abbiamo costruito un modello in grado di prevedere l’ incasso al botteghino di un singolo cinema. Inoltre abbiamo collezionato la stessa tipologia di dati per il mercato americano e applicato il Transfer Learning con l’ obbiettivo di migliorare le performance del modello svedese. Nelle conclusioni abbiamo confrontato le performance dei modelli predittivi allenati usando solo i dati svedesi e i modelli allenati usando dati sia US sia svedesi. Le performance raccolte mostrano che il miglior modello usa solo dati svedesi ed arriva ad una performance di R2=0.83, mentre il miglior modello allenato con transfer Learning arriva ad una performance di R2=0.82. Questi dati mostrano che il Transfer Learning non ha migliorato le performance del nostro modello, almeno nel nostro contesto.
Box office prediction and transfer learning : a study on the US and Swedish market
FOLLONI, ALESSANDRO
2018/2019
Abstract
Motion picture industry has constantly growth in last decade. Still, this is a tricky business and box office failures could happen. Because of this, predicting the success of a movie, a task known as box office forecasting, is a problem worth studying. Forecast the success or flop of movies is a hard task, which involves the study of several variables, like cast, genre, MPAA, social network marketing, interest by the audience and so on. In many cases, such data is hard to collect, and, in many countries, there is a lack of data sources on movies. In this thesis we propose a new idea to solve the problem of scarcity of data in the field of Box Office forecasting. Traditional models use machine learning and statistical analysis to predict the gross of movies. Such studies usually analyze the US market as it is easier to collect data in a big market like America. In this study we propose to use transfer learning. The idea behind transfer learning is the following: use the knowledge from a source domain, which could be the US market, and transfer that knowledge to a target domain. Because of this, the research question that drives this study is the following: Can transfer learning improve the quality of predictions in a situation where few data is available? We focused our study on the Swedish market, which is an example of country where information is hard to find. We collected a data set of Wikipedia page views from the Swedish market and, using machine learning and statistical analysis, we developed a model able to forecast the box office premiere gross of a given movie. We also collected the Wikipedia page views from the US market and applied transfer learning methodologies in order to improve the performance of Swedish models. In transfer learning terms, this means that the US is the source domain and Sweden the target domain. We finally compared the Swedish-only and Swedish-US models to see if transfer learning is beneficial or not. Results show that the best performing model trained only on Swedish data is able to achieve a R2=0.83. The best performing model trained with transfer learning and using both US and Swedish data achieve an R2=0.82. The conclusion of this study is that transfer learning did not improve the performance in our context. These results are only related to this study. We believe that transfer learning was not beneficial to our study as the two studied markets, Sweden and US, behaves differently when it comes to movie selection.| File | Dimensione | Formato | |
|---|---|---|---|
|
Master_Thesis_Polimi_July19.pdf
accessibile in internet per tutti
Dimensione
1.69 MB
Formato
Adobe PDF
|
1.69 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/148515