Recommending Quickly

Daoud Clarke

Hyperparameter Limited

Collaborators

  • Dion Bailey and Tom Pajak (News UK)
  • Carlos Rodriguez (Kainano Ltd.)

Overview

  • Motivation, related work, background
  • Algorithms
    • Real-time content-based recommendations
    • An algorithm for incremental updates to matrix factorization
    • Evaluation
  • Architecture
  • Application to chatbots

Motivation

Personalisation can improve engagement and retention
Need to do this at scale for The Times and The Sun
Needs to be fast
Existing algorithms and systems are not suitable

Related Work

  • Stream-Ranking Matrix Factorization
  • Diaz-Aviles et al. (2012)
  • Users and items must be specified in advance
  • xStreams
  • Siddiqui et al. (2014)
  • New users and items can be specified
  • Not using state-of-the-art matrix factorization algorithms
  • YouTube
  • Covington et al. (2016)
  • Deep learning model using Tensorflow
  • Real-time aspect not described

Algorithms

Experiments

Architecture

Application: chatbots

  • I'd like a pizza.
  • Ok - the aubergine parmigiana again? The portobello mushroom and truffle is also good.
  • Ok, let's try the mushroom.
  • Great. The buffalo mozzarella and smoked tomato salad goes well with that.
  • No thanks.
  • Do you want a Coke to drink like last time? The San Pellegrino Orange is also popular.
  • Just the Coke, thanks.

Conclusion

  • Matrix factorisation is useful
  • Update an existing factorisation using batches
  • Chatbots need personalisation

Thanks!