Upload
janet-minott
View
219
Download
1
Tags:
Embed Size (px)
Citation preview
Adapting Recommendation Diversity to Openness to Experience: A Study of Human BehaviourNava Tintarev, Matt Dennis and Judith Masthoff
University of Aberdeen
UMAP’2013. Rome, Italy
2
Outline
Personality and recommender systems
Experiment – openness to experience and diversity
Results
Limitations
Implications for recommender systems design
UMAP’2013. Rome, Italy
3
Personality traits
Generally it is assumed that: a) traits are relatively stable over time, b) traits differ among individuals (for instance, some people like to
try new things while others prefer to stick to known options), and c) traits influence behaviour (e.g. ordering familiar food at a
restaurant).
Five factor model (Big Five): Extraversion, Agreeableness, Conscientiousness, Emotional Stability, and Openness to Experience
Openness to Experience: active imagination, aesthetic sensitivity, attentiveness to inner feelings, preference for variety, and intellectual curiosity
UMAP’2013. Rome, Italy
4
Personality + Recsys == TRUE?
Recommendation is not all about accuracy.
The tailoring of recommender systems to personality has been found to improve accuracy for sparse data sets and
new users (Hu and Pu 2011), to help predict choices for presidential
candidates (Nunes 2008), to positively impact the acceptance of a
system and recommendations (Hu and Pu 2009, Wu et al 2013)
Openness to experience may make users more receptive to more diverse and potentially serendipitous recommendations.
UMAP’2013. Rome, Italy
5
So what makes for “good” diversity
UMAP’2013. Rome, Italy
6
Diversity != Serendipity
Unexpected and helpful (Ge et al 2010)
Topic diversification approach based on taxonomy-based dissimilarity (Ziegler et al. ,2005). Impacted accuracy negatively.
Re-rank a list of top items was found to improve diversity without a great loss in accuracy (Adomavicius and Kwon, 2011)
Users preferred recommendations from a diversified set of clusters (categorical diversity?), rather than within clusters (thematic diversity?). (Abbassi et al., 2012)
UMAP’2013. Rome, Italy
7
Research Questions
1) there may be a difference in preference for the degree of diversity in recommendations among users,
and
2) within category vs across category diversity in recommendations has not received a great deal of weight in previous literature and would benefit from empirical testing with users.
UMAP’2013. Rome, Italy
8
Experiment: User-as-wizard
UMAP’2013. Rome, Italy
9
Experiment - participants
Amazon Mechanical Turk
120 participants (128 excluded)
57% female, 41% male, 2% undisclosed
Openness to Experience within range for the normal population (TIPI).
Average completion time 5 minutes (up to 30)
UMAP’2013. Rome, Italy
10
Meet Oliver (Dennis et al 2010)
openness_low
Oliver is not interested in abstract ideas, as he has difficulty understanding them. He does not like art, and dislikes going to art galleries. He avoids philosophical discussions. He tends to vote for conservative political candidates. He does not like poetry and rarely looks for a deeper meaning in things. He believes that too much tax money goes to supporting artists. He is not interested in theoretical discussions. Oliver is quite a nice person, and tends to enjoy talking with people.
openness_high
Oliver believes in the importance of art and has a vivid imagination. He tends to vote for liberal political candidates. He enjoys hearing new ideas and thinking about things. He enjoys wild flights of fantasy, getting excited by new ideas.
baseline (no story)
UMAP’2013. Rome, Italy
11
Procedure
Recommend three items (books)
Vary along three dimensions author (0,1)
Same or different genre (0,0.3, 1)
Same, similar or different themes (0,0.3,1)
Almost all themes in common, some themes in common, or no themes in common
Had to justify their choice before moving on to the next recommendation.
UMAP’2013. Rome, Italy
12
UMAP’2013. Rome, Italy
13
Results
UMAP’2013. Rome, Italy
14
Results
No statistically significant effect of story
Effect of order in sequence
But a tendency toward a difference in application of thematic and categorical diversity
UMAP’2013. Rome, Italy
15
Effect of story
Condition DiversityAvg DiversityMax
baseline 1.42 (0.31) 2.00 (0.42)
openness_low 1.41 (0.38) 2.08 (0.51)
openness_high 1.46 (0.30) 2.14 (0.44)
overall 1.44 (0.33) 2.08 (0.46)
• 2-3 things changed!• Slightly higher for openness_high• Difference not reliable – possible ceiling effect• No correlation between the participants (aggregated TIPI
score on) openness to experience and diversity.
UMAP’2013. Rome, Italy
16
Order
Condition Book1 Book2 Book3
baseline 1.04 (0.58) 1.56 (0.49) 1.67 (0.67)
openness_low 1.18 (0.68) 1.35 (0.66) 1.72 (0.74)
openness_high 1.33 (0.63) 1.38 (0.61) 1.68 (0.72)
overall 1.18 (0.64) 1.43 (0.60) 1.69 (0.71)
Book2 >div Book2Book3 >div Book3
p < 0.01 (Bonferroni corrected)
Openness_low starts lowest but `catches up’ by Book3!
UMAP’2013. Rome, Italy
17
Ways of applying diversity
Condition Author Genre Theme
baseline 1.92 (0.70) 1.22 (0.70) 1.12 (0.69)
openness_low 1.83 (0.71) 1.17 (0.63) 1.25 (0.78)
openness_high 1.88 (0.75) 1.45 (0.58) 1.06 (0.49)
overall 1.88 (0.72) 1.28 (0.64) 1.14 (0.66)
• Tendency toward more genre diversity for openness_high• Tendency toward more theme diversity for openness_low• We need to repeat this study!
UMAP’2013. Rome, Italy
18
Results (again)
No statistically significant effect of story
Effect of order in sequence
But a tendency toward a difference in application of thematic and categorical diversity
UMAP’2013. Rome, Italy
19
Limitations
Domain
Is this what people need or what people do?
Predicting openness to experience
UMAP’2013. Rome, Italy
20
Possible implications for recommender systems design
Start narrow go broader
Shift focus toward more thematic diversity (within cluster) for people who are low on openness to experience.
Diversity across genres (across clusters) is still relevant for the majority of users
Worth looking into predicting openness to experience
UMAP’2013. Rome, Italy
21
Wrap-up
We studied the effect of openness to experience on rec diversity
People like to expand each other’s horizons
They start narrow and go wide
Do not really consider personality
But tends toward more thematic diversity for low OE vs more categorical diversity for low OE.
UMAP’2013. Rome, Italy
22
Questions?
This research has been funded by the Engineering and Physical Sciences Research Council (EPSRC, UK), grant ref. EP/J012084/1