Deep Recommender Systems
summary
Overview of practical DL recsys work using a Kaggle ad-click dataset and production tools: trains TensorFlow wide-and-deep and pure deep models, extracts CNN text embeddings and image embeddings, and models sessions with session-based RNNs using negative sampling. Reports that a deep-only model underperforms a linear baseline, wide-and-deep gives modest gains, and a session model improves recall@3 by about 4% over non-ML baselines. Includes implementation notes, training times, and a released dataset and starter code.
AI-written summary · about
chapters (20)
- 0:00Introduction
- 2:11Recommender System Methods
- 6:35Why Deep Learning has a potential for RecSys?
- 8:08The Deep Learning ora of RecSys
- 9:21Research directions in DL-RecSys
- 10:04Learning Item Embeddings
- 12:33Feature extraction from unstructured data
- 16:08Deep Collaborative Filtering
- 18:05Wide & Deep Model example
- 24:21News Recommender Systems
- 25:28Research Objective
- 26:02Conceptual model of factors affecting news relevance
- 26:24Recency factor influence
- 26:54Meta-Architecture Requirements
- 28:25CHAMELEON Meta-Architecture for News RS
- 28:54CHAMELEON - ACR module
- 31:41CHAMELEON - NAR module loss function
- 31:53CHAMELEON - Offline evaluation protocol
- 33:13CHAMELEON - Preliminary Results
- 34:37Next steps
related talks
30mDeep Learning for Recommender Systems
29mFurther enhancements of recommender systems using deep learning
33mDeep Learning at Booking.com
30mLessons from building deep learning recommendation systems
42mDeep Learning Recommender Architecture: Master Thesis to ...
51mTensorFlow Wide & Deep: Advanced Classification the easy way
21mRecsys Keynote: Improving Recommendation Systems & Search in the Age of LLMs
35mDeconstructing Deep Learning