Collaborative Filtering

Project T Final for CS 189/289A: Introduction to Machine Learning @ UC Berkeley

Objectives:

We aim to guide students through the Collaborative Filtering approach to recommendation systems. We want to expose students to 2 paradigms for collaborative filtering: nearest neighbor-style heuristic searches and latent space models that tie into matrix decompositions studied in EECS 16B. The assignment will be a Jupyter Notebook focused on constructing recommendations for a specific dataset -- Netflix title recommendations being a classic example, as well as the MovieLens dataset. We'll start with the former paradigm connecting it to KNN; we will then move on to the models approach, which will address issues with sparsity and show how other techniques -- namely, matrix factorizations similar to Diagonalization/SVD seen in EECS 16AB -- can be used to approach this problem. Lastly, we'll touch upon common approaches for Collaborative Filtering in Industry (Surpriselib, Deep Learning, Regularization), and open problems (e.g. cold start issue).

Specific Objectives:

• Apply knowledge of Pandas and Numpy to load and analyze novel datasets
• Utilize Pandas, Numpy, and Matplotlib to perform Exploratory Data Analysis (EDA) and understand the layout, information, and biases in a given dataset
• Explore cosine similarity as a measurement of likeness between high-dimensional feature vectors of users and movies
• Connect previous ideas of clustering to apply K-Nearest Neighbors towards grouping similar users or movies
• Draw connections to previous matrix factorizations learned in 16AB, and explore how to use gradient descent to learn latent space embeddings
• Use and appreciate packages used in industry (such as Surpriselib) for Collaborative Filtering

Contents

• documentation: Slides and notes on Collaborative Filtering
  • Google Slides Mirror: https://bit.ly/MA_Project_T_Final_Slides
  • Slides Link
  • Notes
• code: Jupyter Notebook assignment and solutions for Collaborative Filtering
  • Student Version
  • Solutions
• assessment: Sample questions to assess student learning
  • Student Version
  • Solutions

Contributors

• Maxwell Chen (@maxhchen)
• Abinav Routhu (@abinavcal)

License