Note

Fredik Nguyen, Sayaka Yamashita, Catalina Regolf, Clementine Lefloch & Miquel Gomez

Predicting Genre From Data – A Movie Classification Project

WEBSITE

https://epfl-ada.github.io/ada-2024-project-alpacadreamagency/

Abstract

The world of movies is extensive and over a century old. The variety of movies that have been made is overwhelmingly large: whether it’s drama, action or romance, horror or sci-fi, there’s a movie out there that caters to each individual’s particular taste. When searching for what to watch, one of the first things we look at is genre — what kind of movie are we in the mood to watch? As such, having a movie tagged with the right genres is crucial for it to find its right audience. This sparked our interest in genre classification: what makes a movie belong to a particular genre? This project will focus on genre and the characteristics of movies that set genres apart. By the end of this project, we want to see if it’s possible to automatically assign genres to movies using their data.

Research questions

Our biggest goal for this project is to make a movie classifier that will assign genres to movies. As such, our research questions revolve around the different characteristics that movies have and their relationships with genre:

• Taking a look at numbers we can ask: Are genre and runtime related? How about the movie budget or its box office revenue? We can ask the same for release dates, movie languages and countries, age ratings, etc.
• Then there’s the stories that movies tell and how they determine their genre. How can we use movie summaries to extract their genres? What about the characters that appear in movies? Is there a trend for the kind of characters that appear in movies for each genre?
• How can we use all this data to build an accurate classifier?

Additional datasets

In order to find more useful features that will help classify movies, we would like to make use of an additional dataset.

The TMDb dataset is a movie dataset that contains information about a million movies. We want to use it to enrich our original dataset with additional information that could be useful for our project like the budget of movies, age rating (whether it’s an adult movie or not), votes, etc. We will merge the datasets so that we only use the movies of the original dataset.

Methods

Data wrangling

We load our databases and merge them by movie title and year so that we keep the movies from the CMU database with the additional information of the TMDb. We also fill in some missing values of the original dataset (runtime, revenue) with values from TMDb. We deal with duplicate movies and clean the data.

The original database made use of over 300 different genres, so we grouped them into 39 new genre categories that the model will use and predict from. In addition we get the hot-vector of the genre of each movie: 1 if the movie has the corresponding genre, 0 o.w.

Data enrichment

From the plot, and the characters we extract additional information that can, potentially, improve our model:

• From the plots we get the sentiment analysis i.e. how positive or negative a movie is. And also related theme words; we get the proportion of words related to “love”, “death” and other topics that can enhance the model predictions.
• From the characters we get how many Male and Female actors there are and their age (grouped into age ranges: 0-20, 20-30…).

Data analysis

We make bar plots, scatter plots and heatmaps to visualize the relationship between movie genres and different movie characteristics. In addition, we make some statistical plots with correlation and violin plots.

Genre Predictor Model

After getting all the data and the new features, we can start creating the model to predict the movie genres. We have in total 5 models.

First of all we two neural network models: a Neural Network with 5 layers, each of its corresponding size for input and output. Also we define some optimizers and the training functions. Hyper parameters are found using a hyper parameter optimizer. A second model with a architecture resembling U-net, is also used.

In addition to the neural network, we have some simpler models to comapare such as ridge regression, k-nearest neighbor, decision tree, random forest.

We feed the model with batches of data in each iteration and use the loss to backpropagate the feedback. After all the iterations we plot the metrics (Precision, Recall and F-score).

Finally, with the trained model, we can test how it does with the “testing set”. The output of which is going to be in a separate file where we can see the movie, its original genre, and the new ones.

As baseline, we use a model that randomly chooses genres.

The best performance was from the ridge regression with a f1-score of 0.42. As for the neural networks they scored at best 0.29. Both models beat randomly choosing genres that f-1 score of 0.12. This indicates that it is possible to predict genres from data.

Proposed timeline and team organization

By the 15th of November we will have established our main goals and ideas, and we will have started working towards them. The next two weeks will be spent reflecting on our project and refining our ideas and methods (and doing Homework 2!).

By the 29th of November we will be ready to spend the following 3 weeks giving it our all to complete our project! We want to work by achieving the following internal milestones:

1. Data cleaning and data analysis (December 6th)

• We will have done some initial data analysis and data cleaning by the Milestone P2 due date, but we might have to develop them further as we carry on working on the project.

2. Movie classifier (December 13th)

• Finishing the classifier will be an important milestone of our project. We will have reached this milestone once we obtain a classifier that works to our satisfaction.

3. Data story (December 16th)

• This milestone will be completed once we obtain the graphics we want to use and write the story we want to tell.

4. Blog post / article (December 19th)

• The final milestone will be putting everything together in the blog post and completing our final hand-in.

5. Milestone P3! (December 20th)

• Final check that everything is in order. Our project will be complete!

These are tentative dates that we will try to follow so that our team achieves a good working rhythm. We will divide our work so that different teammates can work on different milestones and contribute to the completion of the project.

6. Team contribution

• Interactive plots: Fredrik
• Plot summary analysis adn static plots: Sayaka
• ML: Miguel
• Website: Catalina
• Story: Clementine

The rest of the work was a team effort such as data cleaning and analysis

How to use the library

Main file to run is the results.ipynb which runs code contained in scr directory for data processing and ML models. There are some additional files for plots described below

Make sure to download all dependencies needed for the project and all relevant files. Below is a schema of the project structure. The main results of this project can be found on our website: https://epfl-ada.github.io/ada-2024-project-alpacadreamagency/

Project Structure

The data is stored locally as the files are too large. The directory structure of the project looks like below. This schema only contains relevant files in use and that are not in use to build the website.

• assets/: Contains images in addtion to some website relevant assets

  • images/ Contain images
    • ada/ Non generated plots used in the website
    • graphs/ Static plots
    • models/ Plots for models
    • original/ Template images

• data_analysis/: Contains all scripts relevant and result relevant to the webiste

  • output/: Outputs from models and results

  • src/: Contains all files with functions for models and utility used in the project

    • models/ Contains files related to functions used for the model and ml
      • trained_models/ Saved trained models
      • crossvalidation/ Functions for crossvalidation
      • decision_trees/ Functions for decision_trees
      • k_NN/ Functions for k_NN
      • logistic_regression/ Functions for logistic_regression
      • neural_network/ Main file for neural network model
      • regression/ Functions for regression
      • plot_analysis/ Folder for plot analysis amd sentiment analysis
        • sentiment_analysis/ Function for sentiment analysis
        • theme_encoding/ Functions to encode themes into vectors
    • utils/
      • data_utils/ Functions for data loading into notebook and data wrangling
      • general_utils/ Functions for general utilities such as plotting functions
      • evaluation_utils/ Functions for evaluation utils for the model
      • settings.py/ Functions for general utilities such as plotting functions

  • tests/: test for model

  • results/: Main run file for neural network, data handling and data analysis

  • additional_static_plots/: some more static plots

  • interactive_plots/: interactive plots

• gitignore/: List of files ignored by git

• pip_requirements/: File for installing python dependencies

• index.html: Main file for website