Table of Contents

• Content-Based Filtering
• Collaborative Filtering
• Softmax Deep Learning Model

Short Summary: Various recommendation system models.

Content-Based Filtering

• Uses domain expertise and hand-engineered features (i.e., categories for users and items).
  • We can then use similarity functions such as cosine similarity for recommendations.
• Pros:
  • No need for data about other users as recommendations are user-specific (scalable).
  • Can capture the specific user interests and recommend less popular, niche items.
• Cons:
  • Since the features are hand-engineered, a lot of domain knowledge is required.
  • Can only recommend based on existing interests (cannot discover new interests).

Collaborative Filtering

• Uses user embedding \(U \in \mathbb R^{m \times d}\) and item embeddings \(V \in \mathbb R^{n \times d}\) (factors of the matrix).
  • During training, computes \(UV^T\) and uses a loss function to compare to actual results.
• Pros:
  • No domain knowledge necessary since the embeddings are automatically learned.
  • The model can help users discover new interests based on interests of similar users.
  • Embeddings \(U\) and \(V\) are static, and candidates can be pre-computed and stored.
• Cons:
  • Cannot handle fresh items and suffers from the cold-start problem.
  • Difficult to include side features since the embeddings are automatically learned.

Softmax Deep Learning Model

• Uses a deep neural network for predictions.
  • The input is the user query.
  • Output is a probability tensor with size equal to the number of items.
  • Can use the cross-entropy loss for training.
• Pros:
  • Easy to include side features.
  • Easy to handle new queries.
  • Item embeddings \(V\) are static and can be stored.
• Cons:
  • Prone to folding. Need to use techniques such as negative sampling or gravity.
  • Harder to scale to very large corpora (can use hashing, negative sampling, etc).
  • The query embedding needs to be computed at query time (more expensive to serve).