Toward Better Interactions in Recommender Systems: Cycling and Serpentining Approaches for Top-N Item Lists Qian Zhao, F. Maxwell Harper, Joseph A. Konstan (GroupLens) Gediminas Adomavicius (Dept. of Information and Decision Sciences, University of Minnesota) Martijn Willemsen (Human-Technology Interaction group, Eindhoven University of Technology) 1
45
Embed
Toward Better Interactions in Recommender Systems: Cycling and Serpentining Approaches for Top-N Item Lists
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Toward Better Interactions in Recommender Systems:
Cycling and Serpentining Approaches for Top-N Item Lists
Qian Zhao, F. Maxwell Harper, Joseph A. Konstan (GroupLens)Gediminas Adomavicius (Dept. of Information and Decision
Sciences, University of Minnesota)Martijn Willemsen (Human-Technology Interaction group,
Eindhoven University of Technology)
1
Recommender Systems
» Recommender systems typically display the top-N recommended items in order.
2
Example: MovieLens.org Top Picks
3
What’s wrong with this recommender?
4
1st Visit
5
2nd Visit
6
3rd ... Visit
7
8
What’s wrong with this recommender?
9
1st Page
10
5th Page
11
10th … Page
12
13
Excitement
Show Time/Further Exploration
Re-thinking Top-N Recommendation Lists
» Static Top-N in order à the best design for user interaction and temporal experience
» Two missing factors• Fresh vs. stale• Further exploration à worse quality/experience
14
Prior Work: Dynamic/Interactive Recommenders
» Classics• Temporal dynamics (Koren et al. 2010)• CARS (Adomavicius et al. 2011)• Incremental matrix factorization (Luo et al. 2012)
» Interactive (reinforcement) machine learning• Markov decision processes (Shani et al. 2002)• Contextual bandits (Lu et al. 2010)
15
This Work: Cycling and Serpentining
» Cycling demotes items that have been viewed several (3+) times, exposing fresher recommendations.
» Serpentining spreads top recommended items across several pages, offering high-quality items on each page as a user continues to explore.
16
Cycling
Movie Score
M1 5.0
M2 4.9
M3 4.6
M4 4.0
17
Movie Score #display
M1 5.0 3
M2 4.9 3
Cycling
Movie Score
M1 5.0
M2 4.9
M3 4.6
M4 4.0
18
Movie Score #display
M3 4.6 0
M4 4.0 0
M1 5.0 3
M2 4.9 3
SerpentiningMovie Score
M1 5.0
M2 4.9
M3 4.8
M4 4.7
M5 4.6
M6 4.5
M7 4.4
M8 4.3
19
Movie Score
M1 5.0
M3 4.8
M5 4.6
M7 4.4
p. 1
p. 2
p. 1
p. 2
SerpentiningMovie Score
M1 5.0
M2 4.9
M3 4.8
M4 4.7
M5 4.6
M6 4.5
M7 4.4
M8 4.3
20
Movie Score
M1 5.0
M3 4.8
M5 4.6
M7 4.4
M2 4.9
M4 4.7
M6 4.5
M8 4.3
p. 1
p. 2
p. 1
p. 2
When to Cycle
» Within-session: each time when users go back to home page• More perceived change but more confusion?• Reflective of current recommenders in terms of the
change, e.g. Youtube.» Between-session: when users sign in next time• Less disorienting?• But, can users perceive the change?
21
The novel contribution of this work
» Cycling and serpentining approaches are not studied before.• e.g. Youtube is not cycling (Davidson et al. 2010).
» Better understand the effects of dynamic top-N lists on user experience.
22
A between-subjects field experiment on MovieLens.org» Measurements• Objective activity level• Subjective perception (through surveys)
23
Objective Activity Level
» opt out rate (bad experience)» number of page views (negative efficiency or
positive engagement)» number of interested actions. i.e. clicks,
wishlist (positive engagement)» interested rate: number of interested actions
freshness• Higher level of user activities and interested rate
» Between-session cycling or serpentining• Higher level of user activities• Negative effects on subjective perception
40
Efficiency or engagement?
41
Efficiency or engagement?
42
An experience design perspective on recommenders: There is a tradeoff between serving come-and-go users vs. encouraging deeper interaction/engagement !
Messages
» Better understanding of the trade-off between efficiency vs. engagement can help design a better recommender user experience!
» Cycling and serpentining top-N recommendation lists have benefits (higher engagement) but also costs (negative perception)!
» More work combining algorithms and user experience is needed!
43
Thanks! Questions?
» Title: Toward Better Interactions in Recommender Systems: Cycling and Serpentining Approaches for Top-N Item Lists
» Authors: Qian Zhao, Gediminas Adomavicius, F. Maxwell Harper, Martijn Willemsen, Joseph A. Konstan