Neural Networks for Recommender Systems
summary
Neural recommender prototypes use embedding layers in PyTorch with dot‑product scoring, Siamese pairwise losses and an adaptive negative sampler. Code and notebooks on GitHub train on MovieLens 100k and report MRRs: pointwise ≈0.06, pairwise ≈0.08, adaptive ranking ≈0.11. Sequence models compare popularity 0.12, pooling 0.14, LSTM 0.11 and show LSTM wins only on synthetic data with clear sequence patterns. Implementations include CNN image features, LSTM encoders, and working examples so the claimed speed of iteration and loss flexibility is demonstrated.
AI-written summary · about
chapters (12)
- 0:00Hybrid to neural recommendations
- 1:28Neural network advantages
- 5:40Matrix factorization with PyTorch
- 10:03Pointwise loss and evaluation
- 13:57Switching to ranking losses
- 17:05Adaptive ranking and sampling
- 21:28Sequence-based recommendations
- 24:20Preparing sequential data
- 26:10Pooling and RNN representations
- 28:22Testing sequential models
- 29:25Validating with synthetic data
- 30:06Summary of deep learning
related talks
21mNeural Networks for Recommender Systems
57mHands-On with LLM-Powered Recommenders: Hybrid Architectures
29mFurther enhancements of recommender systems using deep learning
01h 21mUnlocking Personalization: A Deep Dive into Modern Recommendation Algorithms
46mBeyond Bag of Words A Practitioner’s Guide to Advanced NLP
42mDeep Learning Recommender Architecture: Master Thesis to ...
30mDeep Learning for Recommender Systems
58mRecommender Systems – A Walkthrough