Slides of my 15 minute presentation on adopting a more user-centric approach in data-driven personalization. Presented during the joint seminar organized by the data-driven decision making (D3M) research theme at School of Business and Economics and the Institute of Data Science (IDS) at Maastricht University.
Seminar Data-Driven Decision Making @ School of Business and Economics & Institute of Data Science
1. Mark Graus
Data-Driven Personalization: Adopting a
User-Centric Approach
Assistant Professor Data Science in Marketing
Department of Marketing and Supply Chain Management, School of Business and Economics
BISS Institute, Heerlen
2. Mark Graus
Academics
• BSc, MSc, PhD @ Eindhoven,
University of Technology
• Human-Technology Interaction
Industry
• Data Scientist Online Customer
Decision Making
• PhD in Collaboration with Philips
6. Problem Statement
Can we show articles for new parents in order
of relevance?
Normal data-driven approaches are available
Normal Complication: Cold-Start
Additional complication: Possible mismatch
between reading and actual interests
Solution: Measure and include psychometrics
in relevance predictions
8. Parenting Styles (Cognitive Part)
• Combined 5 surveys into a new 70-item survey to measure parenting
style.
• Survey considers: Behavior, Cognition and Mother’s Perception of Child.
• SEM Study With 200+ mothers
9. Study Design
(Graus, Willemsen, Snijders, 2018)
Initial Data Collection
(N=181)
• Parenting Styles
• Reading Behavior
Offline Relevance
Predictions
• Baseline
• Reading Behavior
• Parenting Styles
• Both
Revisit the (now
personalized) system
(N=121)
• Reading Behavior
• User Experience
General Top-N
BPRMF
Segment-Based Top-N
BPRMF + User Characteristics
10. Survey-Based Predictions
were Least Accurate
If we were to look at cross-validation,
survey-based performs the worst, hybrid
predictions perform (expectedly) the best.
11. Survey-Based Predictions
Lead to Best User Experience
We measured ‘Perceived level of
Personalization’ and ‘Satisfaction’
Survey-based were the only predictions
that scored significantly higher on both
constructs.
14. 1 A User interacting with a website results in
Behavior and a User Experience.
Based on Behavior we can deduce User
Characteristics
External Data can provide additional
information on User Characteristics
Based on predicted Characteristics, an
Adaptation can be implemented.
The Adaptation will influence Behavior and
User Experience
User Experience can be used to evaluate
Characteristic Predictions and Adaptation
Success
2
2
3
3
4
4
UX
1 1
UX
5 5
5
6
6
6
Personalization – User-Centric Approach
15. Personalization in the Music
Domain
(in collaboration with TU/e and
JADS)
Personalization in Music
• Affective Playlists
• Context-Aware Playlist
• Group Playlists
Role of Engagement
• Different types of listeners
have different needs,
preferences.
17. Different Types of Research Questions
Fundamental
What characteristics of our users are relevant to our goal?
Methodological
How to incorporate characteristics in Predictive Models?
Inferring characteristics from interaction behaviour?
Societal
How does policy (GDPR) affect the possibilities and perception of personalization?
19. Questions/remarks?
Mark Graus
School of Business and Economics, Department of Marketing
and Supply Chain Management
BISS Institute
mp.graus@maastrichtuniversity.nl
twitter.com/newmarrk
medium.com/newmarrk
linkedin.com/in/markgraus
www.markgraus.net
20. References
Graus, M. P., Willemsen, M. C., & Snijders, C. C. (2018, March). Personalizing an Online Parenting
Library: Parenting-Style Surveys Outperform Behavioral Reading-Based Models. In CEUR
Workshop Proceedings (Vol. 2068).
Knijnenburg, B., Willemsen, M., Gantner, Z., Soncu, H., & Newell, C. (2012). Explaining the user
experience of recommender systems. User Modeling and User-Adapted Interaction, 22(4–5),
441–504. https://doi.org/10.1007/s11257-011-9118-4
Editor's Notes
We wanted to measure the effects on both behavior and user experience
Who is a parent?
Parents differ on many aspects, such as how warm they are, how strict they are, how worried they are.
We developed a survey that when answered, tells us what type of parent a person is.
Parents are hard to contact, so we used Facebook ads.
We predicted relevance in four different ways. The reading was used similar to how ratings were used in matrix factorization. The survey data was done by median splits on the attunement and structure dimensions.
This is where data science quite often stops. “We can better predict preferences using reading behavior and surveys.” But we are interested in user experience.
This is already an interesting finding. Subjective experience and objective performance do often not correlate.