Recurrent Neural Networks for Recommendations and Personalization with Nick Pentreath
1 like932 views
The document discusses the use of recurrent neural networks (RNNs) in recommendation systems, highlighting their ability to combine collaborative filtering with deep learning for improved predictions. It covers challenges such as data sparsity and cold start scenarios while presenting future directions for research, including attention models and cross-session recommendations. Overall, it emphasizes the potential of deep learning frameworks to enhance recommendation accuracy and address complex data inputs.
Recurrent Neural Networks for Recommendations and Personalization with Nick Pentreath
- 1. DBG / Oct 3, 2018 / © 2018 IBM Corporation RNNs for Recommendation & Personalization Nick Pentreath Principal Engineer @MLnick
- 2. DBG / Oct 3, 2018 / © 2018 IBM Corporation About @MLnick on Twitter & Github Principal Engineer, IBM CODAIT - Center for Open-Source Data & AI Technologies Machine Learning & AI Apache Spark committer & PMC Author of Machine Learning with Spark Various conferences & meetups
- 3. DBG / Oct 3, 2018 / © 2018 IBM Corporation Center for Open Source Data and AI Technologies CODAIT codait.org CODAIT aims to make AI solutions dramatically easier to create, deploy, and manage in the enterprise Relaunch of the Spark Technology Center (STC) to reflect expanded mission Improving Enterprise AI Lifecycle in Open Source
- 4. DBG / Oct 3, 2018 / © 2018 IBM Corporation Agenda Recommender systems overview Deep learning and RNNs RNNs for recommendations Challenges and future directions
- 5. DBG / Oct 3, 2018 / © 2018 IBM Corporation Recommender Systems
- 6. DBG / Oct 3, 2018 / © 2018 IBM Corporation Users and Items Recommender Systems
- 7. DBG / Oct 3, 2018 / © 2018 IBM Corporation Events Recommender Systems Implicit preference data ▪ Online – page view, click, app interaction ▪ Commerce – cart, purchase, return ▪ Media – preview, watch, listen Explicit preference data ▪ Ratings, reviews Intent ▪ Search query Social ▪ Like, share, follow, unfollow, block
- 8. DBG / Oct 3, 2018 / © 2018 IBM Corporation Context Recommender Systems
- 9. DBG / Oct 3, 2018 / © 2018 IBM Corporation Prediction Recommender Systems Prediction is ranking – Given a user and context, rank the available items in order of likelihood that the user will interact with them Sort items
- 10. DBG / Oct 3, 2018 / © 2018 IBM Corporation Matrix Factorization Recommender Systems The de facto standard model – Represent user ratings as a user-item matrix – Find two smaller matrices (called the factor matrices) that approximate the full matrix – Minimize the reconstruction error (i.e. rating prediction / completion) – Efficient, scalable algorithms • Gradient Descent • Alternating Least Squares (ALS) – Prediction is simple – Can handle implicit data
- 11. DBG / Oct 3, 2018 / © 2018 IBM Corporation Cold Start Recommender Systems New items – No historical interaction data – Typically use baselines (e.g. populariy) or item content New (or unknown) users – Previously unseen or anonymous users have no user profile or historical interactions – Have context data (but possibly very limited) – Cannot directly use collaborative filtering models • Item-similarity for current item • Represent session as aggregation of items • Contextual models can incorporate short-term history
- 12. DBG / Oct 3, 2018 / © 2018 IBM Corporation Deep Learning and Recurrent Neural Networks
- 13. DBG / Oct 3, 2018 / © 2018 IBM Corporation Overview Deep Learning Original theory from 1940s; computer models originated around 1960s; fell out of favor in 1980s/90s Recent resurgence due to – Bigger (and better) data; standard datasets (e.g. ImageNet) – Better hardware (GPUs) – Improvements to algorithms, architectures and optimization Leading to new state-of-the-art results in computer vision (images and video); speech/ text; language translation and more Source: Wikipedia
- 14. DBG / Oct 3, 2018 / © 2018 IBM Corporation Modern Neural Networks Deep Learning Deep (multi-layer) networks Computer vision – Convolution neural networks (CNNs) – Image classification, object detection, segmentation Sequences and time-series – Recurrent neural networks (RNNs) – Machine translation, text generation – LSTMs, GRUs Embeddings – Text, categorical features Deep learning frameworks – Flexibility, computation graphs, auto-differentiation, GPUs Source: Stanford CS231n
- 15. DBG / Oct 3, 2018 / © 2018 IBM Corporation Recurrent Neural Networks Deep Learning Neural Network on Sequences … – … sequence of neural network (layers) – Hidden layers (state) dependent on previous state as well as current input – “memory” of what came before Source: Stanford CS231n – Share weights across all time steps – Training using backpropagation through time (BPTT)
- 16. DBG / Oct 3, 2018 / © 2018 IBM Corporation Recurrent Neural Networks Source: Andrej Karpathy Deep Learning
- 17. DBG / Oct 3, 2018 / © 2018 IBM Corporation Recurrent Neural Networks Deep Learning Issues – Exploding gradients - clip / scale gradients – Vanishing gradients Source: Stanford CS231n Solutions – Truncated BPTT – Restrict sequence length – Cannot encode very long term memory
- 18. DBG / Oct 3, 2018 / © 2018 IBM Corporation Recurrent Neural Networks Deep Learning Long Short Term Memory (LSTM) – Replace simple RNN layer (activation) with a LSTM cell – Cell has 3 gates - Input (i), Forget (f), Output (o) – Activation (g) – Backpropagation depends only on elementwise operations (no matrix operations over W) Gated Recurrent Unit (GRU) – Effectively a simplified version of LSTM – 2 gates instead of 3 - input and forget gate is combined into an update gate. No output gate GRU has fewer parameters, LSTM may be more expressive Source: Stanford CS231n; Hochreiter et al.
- 19. DBG / Oct 3, 2018 / © 2018 IBM Corporation Recurrent Neural Networks Deep Learning Variants – Multi-layer (deep) RNNs – Bi-directional – Deep bi-directional – Attention Source: Stanford CS231n; Denny Britz
- 20. DBG / Oct 3, 2018 / © 2018 IBM Corporation RNNs for Recommendations
- 21. DBG / Oct 3, 2018 / © 2018 IBM Corporation Deep Learning for Recommenders Overview RNNs for Recommendations Most approaches have focused on combining – Performance of collaborative filtering models (especially matrix factorization) • Embeddings with appropriate loss = MF – Power of deep learning for feature extraction • CNNs for image content, audio, etc. • Embeddings for categorical features • Linear models for interactions • RNNs for text Source: Spotify / Sander Dieleman Google Research
- 22. DBG / Oct 3, 2018 / © 2018 IBM Corporation Session-based recommendation RNNs for Recommendations Apply the advances in sequence modeling from deep learning – RNN architectures trained on the sequence of user events in a session (e.g. products viewed, purchased) to predict next item in session – Adjustments for domain • Item encoding (1-of-N, weighted average) • Parallel mini-batch processing • Ranking losses – BPR , TOP1 • Negative item sampling per mini-batch – Report 20-30% accuracy gain over baselines Source: Hidasi, Karatzoglou, Baltrunas, Tikk
- 23. DBG / Oct 3, 2018 / © 2018 IBM Corporation Contextual Session-based models RNNs for Recommendations Add contextual data to the RNN architecture – Context included time, time since last event, event type – Combine context data with input / output layer – Also combine context with the RNN layers – About 3-6% improvement (in Recall@10 metric) over simple RNN baseline – Importantly, model is even better at predicting sales (vs view, add to cart events) and at predicting new / fresh items (vs items the user has already seen) Source: Smirnova, Vasile
- 24. DBG / Oct 3, 2018 / © 2018 IBM Corporation Content and Session-based models RNNs for Recommendations Add content data to the RNN architecture – Parallel RNN (p-RNN) – Follows trend in combining DL architectures for content feature extraction with CF models for interaction data • CNN for image data • BOW for text (alternatives are Word2Vec-style models and RNN language models) – Some training tricks • Alternating – keep one subnet fixed, train other • Residual – subnets trained on residual error • Interleaved – alternating training per mini-batch Source: Hidasi, Quadrana, Karatzoglou, Tikk
- 25. DBG / Oct 3, 2018 / © 2018 IBM Corporation 3D CNNs for Session-based Recommendation RNNs for Recommendations As we’ve seen in text / NLP, CNNs can also be effective in modeling sequences – 3D convolutional models have been applied in video classification – Potentially faster to train, easier to understand – Use character-level encoding of IDs and item features (name, description, categories) • Compact representation • No embedding layer – “ResNet” style architecture – Show improvement over p-RNN Source: Tuan, Phuong
- 26. DBG / Oct 3, 2018 / © 2018 IBM Corporation Challenges Challenges particular to recommendation models – Data size and dimensionality (input & output) • Sampling – Extreme sparsity • Embeddings & compressed representations – Wide variety of specialized settings – Combining session, content, context and preference data – Model serving is difficult – ranking, large number of items, computationally expensive – Metrics – model accuracy and its relation to real- world outcomes and behaviors – Need for standard, open, large-scale, datasets that have time and session data and are content- and context-rich • RecSys 15 Challenge – YouChoose dataset – Evaluation – watch you baselines! • When Recurrent Neural Networks meet the Neighborhood for Session-Based Recommendation Challenges and Future Directions
- 27. DBG / Oct 3, 2018 / © 2018 IBM Corporation Future Directions Challenges and Future Directions Most recent and future directions in research & industry – Improved RNNs • Cross-session models (e.g. Hierarchical RNN) • Further research on contextual models, as well as content and metadata • Attention models: – Attentive Neural Architecture for Music Recommendation – Neural Attentive Session-based Recommendation – Combine sequence and historical models (long- and short-term user profiles) – Domain-specific applications • Contextualized Location Sequence Recommender – RecGAN (yes, GANs and RNNS!) – Applications at scale • Dimensionality reduction, compressed encodings
- 28. DBG / Oct 3, 2018 / © 2018 IBM Corporation Summary Challenges and Future Directions DL for recommendation is just getting started (again) – Huge increase in interest, research papers. Already many new models and approaches – DL approaches have generally yielded incremental % gains • But that can translate to significant $$$ • More pronounced in session-based – Cold start scenarios benefit from multi-modal nature of DL models and explicit modeling of sequences – Flexibility of DL frameworks helps a lot – Benefits from advances in DL for images, video, NLP etc. – Open-source libraries appearing (e.g. Spotlight) – Check out DLRS workshops & tutorials @ RecSys 2016 / 2017, and upcoming in Oct, 2018 – RecSys challenges
- 29. DBG / Oct 3, 2018 / © 2018 IBM Corporation Thank you! codait.org twitter.com/MLnick github.com/MLnick developer.ibm.com FfDL Sign up for IBM Cloud and try Watson Studio! https://ibm.biz/BdYbTY https://datascience.ibm.com/ MAX
- 30. DBG / Oct 3, 2018 / © 2018 IBM Corporation Links & References Wikipedia: Perceptron Stanford CS231n Convolutional Neural Networks for Visual Recognition Stanford CS231n – RNN Slides Recurrent Neural Networks Tutorial The Unreasonable Effectiveness of Recurrent Neural Networks Understanding LSTM Networks Learning Phrase Representations using RNN Encoder- Decoder for Statistical Machine Translation Long short-term memory Attention and Augmented Recurrent Neural Networks
- 31. DBG / Oct 3, 2018 / © 2018 IBM Corporation Links & References Deep Content-based Music Recommendation Google’s Wide and Deep Learning Model Deep Learning for Recommender Systems Workshops @ RecSys Deep Learning for Recommender Systems Tutorial @ RecSys 2017 Session-based Recommendations with Recurrent Neural Networks Recurrent Neural Networks with Top-k Gains for Session- based Recommendations Sequential User-based Recurrent Neural Network Recommendations
- 32. DBG / Oct 3, 2018 / © 2018 IBM Corporation Links & References Personalizing Session-based Recommendations with Hierarchical Recurrent Neural Networks Parallel Recurrent Neural Network Architectures for Feature-rich Session-based Recommendations Contextual Sequence Modeling for Recommendation with Recurrent Neural Networks When Recurrent Neural Networks meet the Neighborhood for Session-Based Recommendation 3D Convolutional Networks for Session-based Recommendation with Content Features Spotlight: Recommendation models in PyTorch RecSys 2015 Challenge – YouChoose Dataset
- 33. DBG / Oct 3, 2018 / © 2018 IBM Corporation