際際滷

際際滷Share a Scribd company logo

Session based recommendations at Wattpad

Mohammad Islam
Mohammad Islam

At Wattpad, we have observed that a majority of the reading activities happen in sequence and they vastly vary by time-of-the-day or namely: user session type. To address this, we designed a recommender system which can model users temporal dynamics and predict what story a user is going to read next based on users sequential behaviour in a particular session.

1 of 11
Download to read offline
Session based Recs At Wattpad Mohammad Islam
Home Feed `
Why Session based ? Wattpad users read different content based on different context. Context examples: Morning/Bed time reads Commute reads Vacation reads etc. A lot of reading activities happen in sequence: Story series, prequels/sequels, serialized stories etc.
Traditional Approach User/Item Vectors = user/item profiles (static) Predicted rating/preference: r = f(u, i)
Example of sessions X1 , X2 , X3 . Xn Y Recommendation problem => Multi Class Classification problem ... ...
Recurrent Neural Network InputSequences Softmax LSTM EmbeddingLayer Output LSTM LSTM ...
Model details Embedding Layer = 128 Dimension Hidden Layers: # of stacked hidden layers = 3 64 hidden units/hidden layer tanh activation dropout Loss: Sparse Categorial Cross Entropy Optimizer: RMSProp
Training Dataset Number of users: 20M Number of stories: 4M Number of sequences: 100M+ Max sequence length: 30 Min sequence length: 3 Minibatch size: 1024 Keras with TF backend Trained on 8-GPU for 2 days
Online Experiment Results Experiment sample size: ~ 1 M users Traffic split: 50/50 Eval metric: normalized mean reading time Experienced Users: 39% New Users (signed up in last 2 weeks): 82%
Learnings Deep Learning opens the door to a whole bunch of new possibilities for recommender systems: Powerful representation Flexibility However, sequential recs is not a solution for every application Intensive pre and post processing required to make it work properly Training is expensive (try with a small dataset with a simple architecture at first) It is highly rewarding (when it works)
References 1. Session-based Recommendations with Recurrent Neural Networks, Bal叩zs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, Domonkos Tikk, ICLR 2016, https://arxiv.org/abs/1511.06939 2. Personalizing Session-based Recommendations with Hierarchical Recurrent Neural Networks, Massimo Quadrana, Alexandros Karatzoglou, Bal叩zs Hidasi, Paolo Cremonesi, RecSys 2017, https://dl.acm.org/citation.cfm?id=3109896 3. Recurrent Recommender Networks, Alex Samola et al, WSDM 2017, https://dl.acm.org/citation.cfm?id=3018689

More Related Content

Session based recommendations at Wattpad

  • 1. Session based Recs At Wattpad Mohammad Islam
  • 3. Why Session based ? Wattpad users read different content based on different context. Context examples: Morning/Bed time reads Commute reads Vacation reads etc. A lot of reading activities happen in sequence: Story series, prequels/sequels, serialized stories etc.
  • 4. Traditional Approach User/Item Vectors = user/item profiles (static) Predicted rating/preference: r = f(u, i)
  • 5. Example of sessions X1 , X2 , X3 . Xn Y Recommendation problem => Multi Class Classification problem ... ...
  • 7. Model details Embedding Layer = 128 Dimension Hidden Layers: # of stacked hidden layers = 3 64 hidden units/hidden layer tanh activation dropout Loss: Sparse Categorial Cross Entropy Optimizer: RMSProp
  • 8. Training Dataset Number of users: 20M Number of stories: 4M Number of sequences: 100M+ Max sequence length: 30 Min sequence length: 3 Minibatch size: 1024 Keras with TF backend Trained on 8-GPU for 2 days
  • 9. Online Experiment Results Experiment sample size: ~ 1 M users Traffic split: 50/50 Eval metric: normalized mean reading time Experienced Users: 39% New Users (signed up in last 2 weeks): 82%
  • 10. Learnings Deep Learning opens the door to a whole bunch of new possibilities for recommender systems: Powerful representation Flexibility However, sequential recs is not a solution for every application Intensive pre and post processing required to make it work properly Training is expensive (try with a small dataset with a simple architecture at first) It is highly rewarding (when it works)
  • 11. References 1. Session-based Recommendations with Recurrent Neural Networks, Bal叩zs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, Domonkos Tikk, ICLR 2016, https://arxiv.org/abs/1511.06939 2. Personalizing Session-based Recommendations with Hierarchical Recurrent Neural Networks, Massimo Quadrana, Alexandros Karatzoglou, Bal叩zs Hidasi, Paolo Cremonesi, RecSys 2017, https://dl.acm.org/citation.cfm?id=3109896 3. Recurrent Recommender Networks, Alex Samola et al, WSDM 2017, https://dl.acm.org/citation.cfm?id=3018689