Transfer Learning Recommender

Overview

Our Transfer Learning Recommender adapts a pretrained bidirectional Variational Autoencoder (BiVAE) to enable cross-domain recommendation. This innovative approach leverages patterns learned from a rich source domain (MovieLens dataset) to make quality recommendations in our target domain (places in Madrid) with limited data.

Technical Implementation

BiVAE Architecture

The Bidirectional Variational Autoencoder consists of:

  1. Encoder Network: Maps input ratings to a probabilistic latent space
  2. Decoder Network: Reconstructs ratings from latent representations
  3. Bidirectional Mapping: Enables transfer between source and target domains
User/Item Inputs → Encoder (μ, σ) → Latent Space (z) → Decoder → Reconstructed Preferences

Cross-Domain Transfer

We bridge the two domains through these key steps:

  1. Pretraining: The BiVAE learns genre embeddings on the MovieLens 100k dataset, which contains rich user-movie rating data

  2. Category Mapping: We establish a semantic alignment between:
    • Movie genres (source domain)
    • Place categories (target domain)

    Examples:

    • “documentary” (movie) → “museum” (place)
    • “nature” (movie) → “park” (place)
    • “food” (movie) → “café” (place)
  3. Embedding Assignment:
    • Each venue inherits its genre’s latent vector from the movie domain
    • Each user’s embedding is computed as the weighted average of embeddings for their top-rated categories

Recommendation Process

The recommendation pipeline follows these steps:

  1. User Embedding Creation:
    • From user category preferences, create a weighted embedding vector
    • Higher weights are given to categories with higher preference ratings
  2. Similarity Scoring:
    • Compute cosine similarity between the user embedding and all place embeddings
    • Apply a category match boost if the predicted category exactly matches the place’s primary tag
    • Apply distance and popularity modifiers
  3. Ranking:
    • Sort places by their final similarity scores
    • Return the top-N recommendations

Implementation in Code

The TransferRecommender class implements this approach:

class TransferRecommender:
    def __init__(self):
        self.base_model = None
        self.genre_embeddings = {}
        self.place_embeddings = {}
        
    def train_base_model(self):
        # Train or load pretrained BiVAE on MovieLens
        # Save genre embeddings to self.genre_embeddings
        # ...
        
    def transfer_to_places(self, places_df):
        # Map place categories to movie genres
        # Assign embeddings to places
        # ...
        
    def get_recommendations(self, user_preferences, user_lat, user_lon, places_df, top_n=5):
        # Generate user embedding from preferences
        # Compute similarities and recommend places
        # ...

Advantages

Limitations and Future Improvements