two-tower-model

🐊 Anime recommendation web app: enter your MyAnimeList username, get a personalized ranked list. Two-stage recommender: two-tower retrieval (PyTorch) + LightGBM ranking.

Building Recommendation Systems in Python and JAX — a fork migrated from TensorFlow/JAX/Flax to pure PyTorch (Python 3.11+). Three sub-projects: Pinterest Shop-the-Look (two-tower CNN, triplet loss), Spotify Million Playlist (embedding-based retrieval, contrastive learning), Wikipedia NLP (GloVe co-occurrence + LSTM text-to-URL).

Transformer-enhanced two-tower recommender on MovieLens-25M with CL-EPIDTN-style contrastive learning for sparse-user and long-tail robustness. FAISS retrieval, neural reranker, FastAPI serving, MLflow experiment tracking, and Ollama-powered recommendation explanations.

An extended ML portfolio evolving from Monash FIT5196 coursework. Features an award-winning EDA report and a Multimodal Sentiment Classifier (RoBERTa + Swin + LoRA) trained on 5.7M Google Maps reviews.

A custom PyTorch Two-Tower Recommendation System for retail platforms. It uses dual neural networks and FAISS vector search for real-time, highly relevant personalized product recommendations, deployed via an interactive Streamlit app.

A multimodal hybrid movie recommendation engine that combines Collaborative Filtering, Content-Based Filtering, and a Hybrid Fusion layer, with genuine multimodal features via OpenAI CLIP (poster images + text metadata).

Movie recommendation app that utilizes machine-learning to recommend new movies.

Interactive RecSys & Time Series Machine Learning portfolio built with Streamlit and Docker

NegSampleLab studies how negative sampling strategy shapes recommender-system learning, ranking quality, false-negative risk, popularity bias, and training efficiency through controlled experiments with random, popularity, in-batch, hard, and mixed negatives.

Two-stage MovieLens recommender — TFX retrieval, LightGBM reranking, FastAPI/Streamlit demo.

Scalable dual-tower retrieval pipeline for movie recommendations, implementing in-batch negative sampling and ultra-fast FAISS vector search.