Anime Recommender Systems Project

By: Muhammed Rüşen Birben, 150220755

Introduction

In this project, I developed a recommender system to suggest animes to users based on their past ratings. My goal was to improve the user experience by providing personalized recommendations. I used python as the programming language.

Dependencies & Installation

The external python libraries used in this project are:

• numpy (1.23.4)
• pandas (1.3.4)
• scikit-learn (1.0.2)
• scipy (1.8.0)
• matplotlib (3.5.0)
• seaborn (0.12.0)
• yake (0.4.8)
• surprise (1.1.3)
• tqdm (4.62.3)

they can be installed via the following commands:

pip install numpy
pip install pandas
pip install scikit-learn
pip install scipy
pip install matplotlib
pip install seaborn
pip install yake
pip install surprise
pip install tqdm

or if you're using linux:

pip3 install numpy
pip3 install pandas
pip3 install scikit-learn
pip3 install scipy
pip3 install matplotlib
pip3 install seaborn
pip3 install yake
pip3 install surprise
pip3 install tqdm

Data

Data is found here on Kaggle which has been prepared by scraping the myanimelist.net webiste. It includes ratings of users and data about the animes like their genres, studios etc. I put the 5 csv that's on the Kaggle dataset inside a folder called data_raw/data/. I created my main .ipynb file in the same directory as data_raw folder.

Methodology

I used both collaborative and content based approach for implementing this recommender system. I utilized user ratings to implement matrix factorization via Stochastic Gradient Descent algorithm for RMSE. I used sparse matrix and extracted user and item representations. And utilized these extracted latent features of users and items to measure similarity. Also using data (meta data) about animes, I found similarities between them. And combining these two approaches I wrote a hybrid function.

Evaluation

I evaluated the performance of my recommender system using several metrics, including RMSE, MAE, and MSE. On the test set, I achieved the following results for ratings between 0 to 1:

• RMSE: 0.289
• MAE: 0.213
• MSE: 0.084

Conclusion

Anime recommender system performed well comparing with similar libraries on python such as Surprise and it was also able to process huge loads of data without causing memory problems unlike Surprise. It was able to provide personalized recommendations to users. In future work, I plan to explore other recommendation algorithms to further improve the system's performance.