Some recommender systems exploit quite complex features and give users the impression that black magic is performed to provide somewhat meaningful recommendations. Users may moreover adopt a rather passive attitude: instead of actively interacting with the recommendations and providing feedback they take the quality of the recommendations for granted. In this talk, we discuss challenges of crowd-based recommender systems and present some strategies for involving people more actively into the flow of computing recommendations.
Carrots for Couch Potatoes: Improving recommendations by motivating the crowd
1. Carrots for Couch Potatoes
Improving recommendations by
motivating the crowd
@fabianabel
2. Definition 1
“Recommender system = black box that knows the answer to
the ultimate question…of life, the universe and everything.”
Hypothesis 1
“The more obscure a recommender system, the higher the
chance that its users are happy with the system.”
3. Definition 2.1
“Data Scientist = folks that can program the smartest
recommender systems.”
Hypothesis 2
“Nobody needs an interaction designer.”
Definition 2.2
“Interaction Designer = folks that think about what users
actually want to do.”
4. Definition 3
“Couch potatoes = users who do not provide input to a
recommender system, but have high expectations towards the
quality of the system.”
Hypothesis 3
“The quality of a recommender system increases with the
number of couch potatoes that are *using* the system.”
5. Goals of Recommender Systems
Make users happy and surprise them with new and
relevant content.
[user perspective]
Deliver content so that monetary success of the
business is maximized.
[business perspective]
6. Problem space
Challenges
• Understanding the users
• Understanding the items
• Coding a good (ensemble
of) recommendation
algorithm(s)
• Evaluation
• Presentation of
recommendations
• …
recommender
system
users
items
recommender system
10. AB test* resultsCTR
Control
group
Group with
Less-Like-This
filtering
-3%
?
*AB tests on XING
- are done in front-end and back-end
components
- typically 50:50 random splits (others:
specific groups; inter-leaving)
- Run for days to weeks significance
level: p-value < 0.01
- Validation includes AA comparison,
BA/BA test, repeating AB test
22. Feedback App
Pros
• Continuous stream of
explicit feedback
• Decoupled from the actual
system
Cons
• Attracts “Haters” more than
fans
• Decoupled from the actual
system
In addition, we also want feedback mechanisms that are
more integrated into the natural interaction flow of the
system.
24. Is Hypothesis 1
wrong as well?
Hypothesis 1
“The more obscure a recommender system, the higher the
chance that its users are happy with the system.”