Further enhancements of recommender systems using deep learning

on YouTube ↗29m

" summary

Overview of concrete methods and results for recommender systems: compares matrix factorization and factorization machines to neural nets, shows where matrix factorization limitations produce wrong neighbors, and argues for end-to-end deep models trained on GPUs using abundant implicit logs. Demonstrates session-focused work—session-based LSTM next-item prediction—and a surprising prototype that ranks well using only movie cover images, plus practical notes on sparse long tails, need for short-session modeling, and when to start with a small alpha product.

AI-written summary · about

" chapters (12)

1. 0:00Start
2. 0:52Recommender systems are considered as an inevitable part of any larger system that is offering some kind of products/services to the final user • Recommender systems are beneficial to both service providers and users • They provide personalized experience
3. 2:56dependent on properties of an item (categorization, description etc.) • independent of other users' behavior • no cold-start problem • cannot recommend anything genuinely "new"
4. 11:56Non-linear Transformation Effective in learning the complex (non-linear) underlying explanatory factors • Sequence Modelling Next item prediction and session based recommendation • Flexibility Plethora of frameworks, good modularization
5. 12:32Interpretability Black-box and limiting explainability • Require large datasets Compared with other domains, maybe not that big of an issue • Hyperparameter Tuning Extensive hyperparameter tuning
6. 14:14Downside to building block methodology • costly preprocessing (keyphrase extraction, topic modelling) • maintaining of each of the modules Newer deep based learning approaches: • Use deep learning as an end-to-end solution
7. 16:59Simple example about possible limitation of MF is caused by the use of a inner product in low dimensional space to describe the more complex user-item relations
8. 20:37Great success of CNNs and RNNs can be credited to the concept of "parameter sharing" It is a neat and efficient way of utilizing the relationship between different input/outputs than a standard neural network
9. 20:57The sequential structure of session or click-logs are highly suitable for interpretation provided by the recurrent neural network models (RNN). Real-life recommender systems often rely on short session-based data rather than long user activity history (e.g. Youtube, Netflix etc.)
10. 24:23Basic RNN highly suitable for sequence modeling but suffer from short-term memory problem LSTM (long short-term memory) and GRU (Gated recurrent unit) are created in order to mitigate the issue
11. 25:48Many real-world recommendation problems are best solved through thoughtful analysis, feature engineering, simple models and effective feedback systems
12. 26:27Product sense • Models without product impact are of no use. • Think how your work influences product decisions and solves people problems . Always be involved with your final product

" related talks

• 21mNeural Networks for Recommender SystemsOlivier Grisel · dotAI 2017
• 32mLet's talk about Recommendation Systems!Mihailo Joksimovic · DSC Europe 2022
• 30mDeep Learning for Recommender SystemsAlexandros Karatzoglou · WeAreDevelopers AI Congress Vienna 2018
• 38mRecommender systems with TensorflowGuillaume Allain · PyData London 2017
• 01h 31mHands on - Build a Recommender systemCamille Couturier · PyData Amsterdam 2019
• 01h 21mUnlocking Personalization: A Deep Dive into Modern Recommendation AlgorithmsSarang Gupta · DSC Europe 24
• 20mBuilding Recommender systemSwaroop Krothapalli · The Fifth Elephant 2015
• 37mA primer on recommendation systems.Manuel Garrido Peña · PyData Madrid 2016