Recommender systems are the secret ingredient behind personalized online experiences and powerful decision-support tools across various industries, including retail, entertainment, healthcare, and finance.
These systems work by understanding the preferences, past decisions, and characteristics of many people. For instance, a recommender can predict the types of movies an individual will enjoy based on the movies they’ve previously watched.
Recommender systems manage three key objects: users, items, and user-item interactions. These objects are interconnected and influence each other through various relations. Therefore, recommender systems can be effectively modeled using graphs, which capture the complex and heterogeneous nature of the available data. It's no surprise that in recent years, integrating graphs into recommender systems has gained significant attention from researchers and practitioners.
In a graph, nodes represent entities (such as users and items), while edges represent the relationships between these entities. Mapping entities and their attributes into a graph helps in understanding their mutual relations.
To leverage this structure, Graph Neural Networks (GNNs) will be applied. GNNs have recently become very popular due to their ability to learn complex relationships or interactions across various problems. They are among the best-performing architectures for graph learning tasks. The key idea in GNNs is to iteratively aggregate feature information from local graph neighborhoods using neural networks. This aggregation step allows each node to learn a more general representation from its local neighborhood.
The goal of this project was to build a recommender system opon the MovieLens 100K dataset to predict how a user would rate a certain movie. The data has been taken from Neo4J's sandbox project 'Recommendation'. This is a enriched dataset with many useful movie features. A preprocessed version of this dataset is stored in ./data
A detailed report about the project as well as the presantation slides can be found in ./docs
In this project I worked with a tech stack including:
- Neo4j as database (raw data)
- Hugging face to embed movie titles an plots
- PyTorch and PyTorch Geometric (PyG) to build and train a GNN based recommender system
- Python data science stack (pandas, numpy etc.) for data processing
Apart from PyTorch Geometric all libraries can be installed with pip. I recommend to create a virtual environment with Python 3.9. The requirements.txt can be found in the project directory. Code snippets to install PyTorch Geometric can be found in 00_Install_PyTorchGeom.ipynb.
It was a very interesting project. I learned a lot about Graph Neural Networks and Recommender Systems. Although, the model performed very well, there are some points I think could be improved:
- Normalize user ratings: Some users tend to rate movies higher than others. The ratings should be normalized for each users, so that each user has the same mean rating.
- Lower k: For precision and recall at k. I used k=20, which is too high and leads to wrong conclusions. On average there are less than 20 samples of ratings above 3.5 for each user in the testset. Therefore, precision@20 is misleading. For a more accurate precision@k score, k should be lowerd to 5.
- My Recommender system does not consider cold starts: With the dataset used in this project I did not have the problem of cold starts. Cold start is a well known problem with recommender systems. I describes the problem that, a recommender system would never recommend a newly released movie to an user because there was simply no interaction with this movie during training. There are different ways to handle the problem of cold start. One is binning - the recommender systems does not recommend items but rather bins of items. A postprocessing step (e.g. business rule) than extracts a single item from the recommended bin to the user. Binning leads also to a reduction of the item space, which could be beneficial if the item space is very large.