5. 5
Can we generate personalized natural
language explanations at scale?
Background Introduction Method Experiment Result Discussion
6. 6
Tintarev et. al. (2012) Evaluating the effectiveness of explanations
for recommender systems.
Vig. et. al (2008) Tagsplanations.
Background Introduction Method Experiment Result Discussion
7. 7
Can we do better than formulaic
explanations?
Background Introduction Method Experiment Result Discussion
10. 10
From your MovieLens profile it seems that you
prefer movies tagged as visual, Gravity is unlike
what we have seen on a cinema screen before and
arguably it has one of the best uses of 3D in a movie.
Background Introduction Method Experiment Result Discussion
11. 11
From your MovieLens profile it seems that you
prefer movies tagged as intense, the movie is a
pretty intense ninety minutes, with Bullock's
character constantly battling one catastrophe after
another, and all of it is amazing to see.
Background Introduction Method Experiment Result Discussion
14. 14
From your MovieLens profile it seems that you prefer movies
tagged as [Topical Dimension], [Natural Language Explanation].
1. Model topics of items
2. Generate explanations
3. Model user interest and present matching
explanation
Background Introduction Method Experiment Result Discussion
15. 15
Model topics of items
1
Background Introduction Method Experiment Result Discussion
Crowd recruited from Amazon Mechanical Turk
16. 16
Top 20 most relevant tags
Background Introduction Method Experiment Result Discussion
mafia, gangster, gangsters,
mob, crime, mentor
violent, narrated, violence, stylish,
visceral, stylized, bloody, brutality
masterpiece, storytelling, drama,
dialogue
interesting, original
17. 17
Build semantic similarity
graph (word2vec trained on
IMDB reviews)
Clustering
(Affinity
Propagation)
Input to crowd
Background Introduction Method Experiment Result Discussion
28. 28
• Generate natural language explanations
for 100 movies with $3.90/movie
• Survey 216 MovieLens users
Background Introduction Method Experiment Result Discussion
29. 29
• Within subject design
• Two random unseen movies
• One with baseline and another with
natural language explanation
Background Introduction Method Experiment Result Discussion
30. 30
• Baseline: user preferred relevant topics
We recommend the movie because you like the
following features: [tag1, ..., tag5]
Vig. et. al (2008) Tagsplanations.
Background Introduction Method Experiment Result Discussion
34. 34
Users are have more trust in
natural language explanations
2
Background Introduction Method Experiment Result Discussion
35. 35
15%
17%
66%
57%
19%
26%
24%
22%
48%
42%
28%
36%
I trust the explanation.
The explanation reflects my preferences about this movie.
CROWD
TAG
CROWD
TAG
100 50 0 50 100
Percentage
Strongly disagree Disagree Neutral Agree Strongly agree
Background Introduction Method Experiment Result Discussion
36. 36
Users perceive natural language
explanations to contain more
appropriate amount of information
3
Background Introduction Method Experiment Result Discussion
37. 37
34%
55%
45%
19%
21%
26%
The explanation contains right amount of information.
CROWD
TAG
100 50 0 50 100
Percentage
Strongly disagree Disagree Neutral Agree Strongly agree
4%
6%
63%
55%
32%
39%
Changes in response regarding knowledge about a movie
CROWD
TAG
100 50 0 50 100
Percentage
Response −2 −1 0 1 2 3 4
Background Introduction Method Experiment Result Discussion
38. 38
4
Little difference in helping users
make decision
Background Introduction Method Experiment Result Discussion
45. 45
• Mixed computation approach
• Human effort to
- Refine item topic clusters
- Synthesize review quotes
• Better user experience than tag based
explanations
Background Introduction Method Experiment Result Discussion
46. 46
Questions?
Shuo (Steven) Chang @ Quora
schang@cs.umn.edu
http://www-users.cs.umn.edu/~schang
Crowd-Based Personalized Natural Language Explanations for
Recommendations
47. 47Intro New user Model topics Recommend Explain Discussion
mafia, gangster, gangsters,
mob, crime, mentor
violent, narrated, violence, stylish,
visceral, stylized, bloody, brutality
masterpiece, storytelling, drama,
dialogue
interesting, original
Quotes about
drama masterpiece, story-
telling, dialogue :
As much as the true events of Henry’s life have more
than likely been dramatised and glamourised to a
certain extent, the essence of this film IMO is that it is
still a brilliantly damning portrayal of the characters
and lifestyle of mobsters.
The consistently fine acting by the large ensemble cast
(both known and unknown), the cinematography,
editing, dialogue, brilliant use of music, it’s all
breathtaking.
The dialogue is incredible.
Storytelling with impeccable pacing, this is what it’s
like when a master composer conducts his masterpiece.
If ever the word ‘masterpiece’was meant to be used, it
was for this film. ‘Goodfellas’is a masterpiece, pure
and simple.