This document provides an overview of Elsevier's work using machine learning techniques like collaborative filtering and learning to rank to improve article recommendations for researchers. It discusses using collaborative filtering on browsing logs to initially recommend related articles, then using learning to rank to re-rank those recommendations higher based on features like reputation, topics, citations and user engagement data. Evaluation shows the learning to rank approach improved user engagement by 9-10% over collaborative filtering alone. Future work may explore alternative approaches like graph-based methods or deep learning.
5. | 4
Our Users
We combine content and data with analytics and technology to help:
RESEARCHERS
to make new discoveries and
have more impact on society
CLINICIANS
to treat patients better
and save more lives
NURSES
throughout their careers
and to help save lives
6. | 5
Researcher’s Journey
Help me
stay up to
date
Help me
showcase my
work
Help me
organise my
writing
Help me
make peer review more
rewarding
Help me
publish faster
Help me
manage research
data
Help me
with my editorial decisionsHelp me
connect with the right
people
Help me
secure funding
Help me
read and evaluate
articles
7. | 6
Being the best researcher you can be!
• Good researchers are on top of their game
• Large amount of research produced
• Takes time to get what you need
• Help researchers by recommending relevant content
13. | 12
ScienceDirect – related article recommender
• Scientific publication
database
• 15 million articles
• 14 million monthly visitors
14. | 13
Science Direct V1 Recommender
• Goal
- Present users with related articles based on the article they are
reading
• Start simple & iterate
- Browsing logs to generate item-to-item CF recommendations
- Article content as business logic filtering based on recency, article
type
15. | 14
Item-based kNN Collaborative Filtering
Recommend articles that are similar to the ones you browsed
- Similarity is based on article co-occurrences in users’ browsing sessions
- “Users who read x also read y”
Identify similar articles using cosine similarity: cos $%, $' =
)*×),
)* × ),
Why we use it?
- Gives good results
- Scales relatively well
- Relatively simple to implement
16. | 15
Evaluation: Session prediction task
• Article browsing logs:
• Predict what users would browse next
• Time-split evaluation
< "#""$%&'(, *+,$-.#'(, *--#""/$0# >
Train model Query
Ground
truth
Time, user interactions
Test
17. | 16
CF & Significance Weighting
• Scale down cosine similarity with significance weighting
• Preference is given to high co-occurrence neighbors
- k – min # sessions in common to get original cosine similarity
• Alternative – minimum co-occurrence threshold
- Significantly reduces the catalogue coverage
score &', &) = min 1,
|0'⋂0)|
2
x 345678(&', &))
18. | 17
Other CF Improvements
• Min/max filters for # articles per user-session & # users-sessions per article
• ~ 12 months of browsing logs
- gives good coverage
- removes cyclical nature of academic year
- focuses on more “current” interactions
• Bias for recent activity using time decay functions (e.g. exponential)
• Using article content as business logic filters for recency and article types
21. | 20
A Wealth of Data
• Usage Data
- Logged-in activity
- Alt-metrics,
popularity, trending
• Social Features
- User profiles
- Social network
- Collaboration groups
• > 60 million records: journals, conferences,
books, patents …
• The most accurate and complete citation & co-
author graphs
• Reputation metrics for articles, authors and
journals
• > 15 million full text articles
• Article browsing logs
• Recommender impression and click logs!!!
22. | 21
Learning to Rank (LtR)
CF
candidates
Enriched
candidates
Re-ranked recs
Features
LtR
model
Use CF as candidate selection
Enrich with item and user features
Re-rank results based on learnt model optimised for CtR
23. | 22
LtR Features
Reputation &
Alt-Metrics Text
Topics
Temporal
Images: wsj, alamy, bookedelic
CF similarity
score &', &) = min 1,
|0'⋂0)|
2
x 345678(&', &))
Citation Network
24. | 23
LtR Models
• Set of labelled query documents and their associated recommended
documents with feature vectors and relevance judgements
• Different optimization objectives – point-wise, pair-wise & list-wise
• RankLib java-based LtR package
- RankNet – pair-wise neural network algorithm
- LambdaRank – extension of RankNet optimizing list-wise IR metrics such as
NDCG
- LambdaMART – list-wise approach combining LambdaRank and MART
< "#$%&'()*, %$)'(),*-ℎ/$0-#%$12, %$34)(%$*2 >
25. | 24
Recommender Logs
LtR requires labelled training data that represents user preferences
in relation to the recommendation lists
Recommender Logs
- Impressions – recs shown to the user
- Clicks & conversions – recs the user engaged with
- Timestamp – when the event happened
- Page-load ID – groups recs that were shown at the same time
26. | 25
Training data for LtR Models
• Query-recommendations pairs with relevance labels inferred from
recommender logs
• For each query article
- Aggregate the recommended articles across all user sessions
- Count # impressions & clicks for each recommendation
- Compute graded relevance scores based on CTR
27. | 26
Explore/Exploit via Dithering
Slightly shuffle the list of recommendations
• Allows for the exploration of the list
• Gives the impression of freshness
• Reduces some of the bias in LtR training data
!"#$%&'()*+*& = log $012 + 4 0, log 7
where < =
∆ $012
$012
and tipically < ∈ [1.5,2]
28. | 27
Evaluation: Click prediction task
• Data:
• Rank higher the recommendations users engage with
• Time-split evaluation
< "#$%&'(&)$*+%,-, &%*(&)$*+%/$)ℎ1%2)#&%3, &%+425%+ >
Train model
Validation
Set
Test
Set
Time, user interactions
29. | 28
Results
• LtR improved the quality of recommendations
- 9-10% improvement in user engagement
- Winner is LambdaMART - GBDT with list-wise optimization
• LtR increased journal diversity in recommendation lists
• LtR promotes recently published articles in the last year
• Best ranking model combines usage data with rich structured
network and meta data
30. | 29
Offline evaluation should match the online challenge
• Candidate generation – Collaborative Filtering – session prediction task
• Re-ranking candidates – Learning-to-Rank – click prediction task
31. | 30
LtR in Production
LtR
rescoringIBCF
Recommendation
clicks
Training data
LtR
model
Article
views/downloads
33. | 32
Alternative Approaches
Graph-based approaches
- Random walks for candidate generation
Deep Learning
- Learn more complex features for LtR
- Neural embeddings for candidate
generation
- Hybrid systems for ranking
34. | 33
Evaluation – correcting for bias & confounding
• Algorithm confounding
- How algorithmic confounding in recommendation systems increases homogeneity and
decreases utility. Allison J. B. Chaney, Brandon M. Stewart, and Barbara E. Engelhardt
(RecSys '18).
• Explore/exploit – multi-armed bandits
- Explore, exploit, and explain: personalizing explainable recommendations with bandits.
James McInerney, et al. (RecSys ‘18).
• Counterfactuals
- Counterfactual reasoning and learning systems: The example of computational advertising.
Bottou, Léon, et al. (JMLR 2013).
37. | 36
Recommender Team Publications
Hristakeva, M., Kershaw, D., Pettit, B., Vargas, S., & Jack, K. (2019). Academic recommendations:
The Mendeley case. In Collaborative Recommendations: Algorithms, Practical Challenges and
Applications.
Pettit, B., Hristakeva, M., Kershaw, D. & Jack, K. (2018). Learning to Rank Research Articles: A case
study of collaborative filtering and learning to rank in Science Direct.
Hristakeva, M., Kershaw, D., Rossetti, M., Knoth, P., Pettit, B., Vargas, S., & Jack, K. (2017). Building
recommender systems for scholarly information. WSDM2017.
Rossetti, M., Vargas, S., Pettit, B., Kershaw, D., Hristakeva, M., & Jack, K. (2017). Effectively
identifying users’ research interests for scholarly reference management and discovery. WSDM2017.
Vargas, S., Hristakeva, M., & Jack, K. (2016). Mendeley: Recommendations for
Researchers. RecSys ’16
38. | 37
References
From RankNet to LambdaRank to LambdaMART: An Overview (2010). Christopher J. C. Burges.
On Application of Learning to Rank for E-Commerce Search by Shubhra Kanti Karmaker Santu,
Parikshit Sondhi, and ChengXiang Zhai (SIGIR 2017).
Recommender Systems Handbook (2010). Francesco Ricci, Lior Rokach, Bracha Shapira, and Paul
B. Kantor.
Practical Machine Learning: Innovations in Recommendation (2014).
Ted Dunning and Ellen Friedman. O'Reilly Media, Inc.
Pixie: A System for Recommending 3+ Billion Items to 200+ Million Users in Real-Time by Chantat
Eksombatchai, Pranav Jindal, Jerry Zitao Liu, Yuchen Liu, Rahul Sharma, Charles Sugnet, Mark
Ulrich, and Jure Leskovec (WWW 2018).
Getting Deep Recommenders Fit: Bloom Embeddings for Sparse Binary Input/Output Networks by
Joan Serrà and Alexandros Karatzoglou (RecSys 2017)
39. We're hiring, come speak to
us!
https://www.elsevier.com/about/careers/technology-careers