Speaker: Oliver Gindele, Data Scientist at Datatonic Title: Recommender systems with TensorFlow Abstract: Recommender systems are widely used by e-commerce and services companies worldwide to provide the most relevant items to the user. Many different algorithm and models exist to tackle the problem of finding the best product in a huge library of items for every user. In this talk, Oliver explains how some of these models can be implemented in TensorFlow, starting from a collaborative filtering approach and extending that to deep recommender systems. Speaker Bio: Oliver is a Data Scientist at Datatonic with a background in computational physics and high performance computing. He is a machine learning practitioner who recently started exploring the world of deep learning. Thanks to all TensorFlow London meetup organisers and supporters: Seldon.io Altoros Rewired Google Developers Rise London

1. Recommender Systems with Tensorflow
Oliver Gindele
@tinyoli
oliver.gindele@datatonic.com
14.02.2018

2. Who is Oliver?
+ Data Scientist
+ PhD in computational physics
Who is datatonic?
We are a strong team of data scientists, machine learning
experts, software engineers and mathematicians.
Our mission is to provide tailor-made systems to help your
organization get smart actionable insights from large data
volumes.

3. Recommender Systems

4. Recommender Systems

5. Collaborative Filtering - Introduction

6. Collaborative Filtering - Introduction
Objective:
+ Netflix price (2009)
+ Solve via SVD (ALS or SGD)
+ Regression problem

7. Finding Love with Numbers

8. Online Dating Dataset - LibimSeTi

9. Online Dating Dataset - LibimSeTi
+ http://www.libimseti.cz/
+ 2005
+ 17,359,346 ratings
+ 135,359 users
+ Ratings: 1-10
+ Female (%): 69
+ Male (%): 31
+ Mean(rating): 5.9
+ Std(rating): 3.1
userId profileId rating gender

10. Tensorflow: High Level APIs
Dataset:
Estimator:

11. Dataset API (tf.data)

12. Estimator API (tf.estimator)

13. MF Model

14. Going Deeper - Beyond MF
Neural Collaborative Filtering (He et al.)

15. Going Deeper - Beyond MF
User Metadata Item Metadata
userid age gender
1225 ‘30-40’ ‘F’
itemid genre length
44044 ‘Drama’ 129
Add user and item (profile)
characteristics

16. Results - LibimSeTi
MF MLP MF + MLP Research [1]
RMSE 2.137 2.112 2.071 2.077
MAE 1.552 1.541 1.432 1.410
[1] Trust-Based Recommendation: an Empirical Analysis, O’Doherty, Jouili, Van Roy (2008)
Training details:
+ 40 epochs
+ MLP: 4 layers (256 units pyramid)
+ Adam optimiser
+ Results calculated on held out test set (5 rating per user)
+ No tuning of hyperparameters

17. Further Approaches

18. Recommender Systems - Improvements
+ Implicit feedback (Hu, Koren, Volinsky 2008):
+ Logistic Matrix Factorisation (Johnson, Spotify, 2014):
+ Negative Sampling
+ Ranking metrics for better evaluation: HitRate@K, NGCG@K

19. Deep Recommender Systems - Advances
Wide & Deep model (Cheng et al. 2016)
Continuous Features Categorical Features

20. Deep Recommender Systems - Advances
Deep Neural Networks for YouTube Recommendations (Covington, Adams, Sargi 2006)

21. Takeaways
+ Tensorflow can do more than vision or translation
+ High level APIs make model building and training painless
+ Custom algorithms and specific loss functions are easily implemented
+ Embeddings and hidden layers allow for many ways to improve a recommender system

22. Thank you.
Blog.datatonic.com
facebook.com/datatonic
linkedin.com/company/datatonic
twitter.com/teamdatatonic