A tensorflow recommending system for news — Fabrício Vargas Matos (Hearst tv) @PAPIs Connect — São Paulo 2017

News recommendations are particularly challenging given the high number of new contents produced every day and the fast deterioration of its value for the users, demanding models and infrastructure able to deal with those nuances and serve a newly trained model about 100 times per day. Attending this presentation you're going to follow a detailed overview of how R&D team of Hearst's TV division is putting together Google BigQuery, Kubernetes cluster and Tensorflow to build a hybrid recommendation system combining model-based matrix factorization, content recency, and content semantics through NLP.

1. A Tensorflow
Recommending System
for News
Fabricio Vargas Matos

2. Manhattan, NYTV Stations
Local and National News

3. Article’s page: recommendations
for continuous scroll section
Recommended articles

5. Agenda
1.Recency and cold-start problem
2.Data acquisition
3.Matrix factorization
4.Tensorflow implementation
5.Hybrid Model: NLP and feature engineering
6.Hybrid Model: Hybrid matrix factorization
7.Conclusions

8. Cold-start problem
Existent
Items
New
Items
Existent Users New Users

9. Cold-start solution
Existent
Items
New
Items
Existent Users New Users
Not personalized!
Curated by Editors
+
Highly viewed

10. Cold-start solution
Existent
Items
New
Items
Existent Users New Users
Not personalized!
Curated by Editors
+
Highly viewed
Hybrid
Matrix
Factorization

11. Data Acquisition
Page views with
user’s time on page
Google AnalyticsGoogle BigQuery CMS
Content corpus: title,
body, timestamp,
meta-data (sections,
tags, etc.)
Contents
TFRecord/CSV files

12. "Users x Items" Sparsity
Dataset Sparsity
MovieLens (movies) 98.61%
Netflix (movies) 98.82%
TV Stations (news) 99.94%
Yahoo! KDD (music) 99.96%

13. Matrix Factorization

14. VU
Latent Factors Model
R
Items
Users
≈
Latent
factors
Latent
factors
Items
x
user bias item bias
i
j
i
j
R[i,j] ≈ U[i] x V[j]

15. TF code: factorization op
(…)

16. TF code: train op

17. Initial Results
•Training time ≈ 15min (Kubernetes cluster)
•TimeOnPage Prediction Error (RMSE) ≈ 125 sec
•Qualitative recommendation tests with chosen
‘personas’ revealed poor personalization

18. Hybrid Matrix
Factorization Model

19. Natural Language
Processing
Concatenate content data
(title, body, sections, tags, …)
Remove stop words, symbols
and HTML tags
Train word2vec Neural Network
Combine all word-vectors of
each article into one (doc2vec)
CMS
articles
doc2vec
contents

20. Contents Data
Visualization

22. Entertainment
National News
Health
Sports
Local News

45. Features Engineering
NLP (doc2vec)
items clustering (k-means)
embed items:
similarity to each cluster centroid
embed users:
viewed contents combined
CMS
articles
k-dimension
items/users
embeddings
Google
Cloud
Storage

50. Items Parallel coordinates: 40 features/clusters

51. Feature #1: Similarity to
cluster #1

52. Feature #39

53. Who are they?
Magenta contents (health) with high
values for feature #1 (economy)?

55. Content/User Embeddings
+
Matrix Factorization

56. VU
Matrix Factorization
R
Items
Users
≈
Latent
factors
Latent
factors
Items
x
user bias item bias
i
j
i
j
R[i,j] ≈ U[i] x V[j]

57. Hybrid Matrix Factorization
•R ≈ U
*
x V
*
where:
•U
*
= UUsersxKClusters x AKClustersxLatent_factors
•V
*
= BLatent_factorsxKClusters x VKClustersxItems
*Only A and B are variables to be trained. U and V are constants.

58. TF code: factorization
Now:

59. Results
•Training time ≈ 20min (Kubernetes cluster)
•TimeOnPage Prediction Error (RMSE) ≈ 100 sec
(20% better)
•Qualitative recommendation tests with chosen
‘personas’ revealed very good personalization
•R&D Project - Not yet publicly available

60. Let’s talk online
fabriciovargasmatos@
Fabricio Vargas Matos