Context-Aware Recommender Systems are advisory applications that exploit users’ preference knowledge contained in datasets of context-dependent user ratings, i.e., ratings augmented with the description of the contextual situation detected when the user experienced the item and rated it. Since the space of context-dependent ratings increases exponentially in size with the number of contextual factors, and because certain contextual information is still hard to acquire automatically (e.g., the user’s mood or the travellers’ group composition), it is fundamental to identify and acquire only those factors that truly influence the user preferences and consequently the ratings and the recommendations. In this paper, we propose a novel method that estimates the impact of a contextual factor on rating predictions and adaptively elicits from the users only the relevant ones. Our experimental evaluation, on two travel-related datasets, shows that our method compares favorably to other state-of-the-art context selection methods.
Contextual Information Elicitation in Travel Recommender Systems
1. ENTER 2016 Research Track Slide Number 1
Contextual Information Elicitation in
Travel Recommender Systems
Matthias Braunhofer and Francesco Ricci
Free University of Bozen - Bolzano, Italy
{mbraunhofer,fricci}@unibz.it
http://www.inf.unibz.it
2. ENTER 2016 Research Track Slide Number 2
Agenda
• Introduction
• Related Work
• Selective Context Acquisition
• Experimental Evaluation and Results
• Conclusions
3. ENTER 2016 Research Track Slide Number 4
Context-Aware
Recommender Systems (CARSs)
• CARSs provide better recommendations by
incorporating contextual information (e.g., time
and weather) into the recommendation process
STS (South Tyrol Suggests)
4. ENTER 2016 Research Track Slide Number 5
Context Acquisition
Problem of CARSs
• How to identify and acquire the truly relevant
contextual factors that influence the user
preferences and decision making process?
5. ENTER 2016 Research Track Slide Number 6
STS w/o Selective
Context Acquisition
We can’t elicit the conditions for all the available
contextual factors when the user rates a POI.
6. ENTER 2016 Research Track Slide Number 7
STS w/ Selective
Context Acquisition
Rather, we must elicit the conditions of a small
subset of most important contextual factors.
8. ENTER 2016 Research Track Slide Number 9
Context Acquisition
Problem in Commercial Systems
• Numerous commercial systems in the tourism
domain face the context acquisition problem
TripAdvisor Foursquare
9. ENTER 2016 Research Track Slide Number 10
A Priori Context Selection
• Web survey in which
users evaluate the
influence of contextual
conditions on POI
categories
• Allows to identify the
relevant factors before
collecting ratings
(Baltrunas et al., 2012)
10. ENTER 2016 Research Track Slide Number 11
A Posteriori Context Selection
• Several statistic-based
methods for detecting
the relevant context
after collecting ratings
• Results show a
significant difference in
prediction of ratings in
relevant vs. irrelevant
context (Odić et al., 2013)
12. ENTER 2016 Research Track Slide Number 13
Parsimonious and
Adaptive Context Acquisition
• Main idea: for each user-item pair (u, i), identify
the contextual factors that when acquired with
u’s rating for i improve most the long term
performance of the recommender
– Heuristic: acquire the contextual factors that have the
largest impact on rating prediction
• Challenge: how to quantify these impacts?
13. ENTER 2016 Research Track Slide Number 14
CARS Prediction Model
• We use a new variant of Context-Aware Matrix
Factorization (CAMF) (Baltrunas et al., 2011) that treats
contextual conditions similarly to either item or
user attributes
ˆruic1,...,ck
= (qi + xa )T
(pu + yb )+ ri
b∈A(u)∪C(u)
∑ + bu
a∈A(i)∪C(i)
∑
Latent vector of item i Latent vector of user u
Latent vectors of
conventional (e.g., genre)
and contextual item
attributes (e.g., weather)
Avg. rating for item i
Bias of user uLatent vectors of
conventional (e.g., age)
and contextual user
attributes (e.g., mood)
14. ENTER 2016 Research Track Slide Number 15
Largest Deviation
• Given (u, i), it computes a relevance score for
each contextual factor Cj by first measuring the
“impact” of each contextual condition cj C∈ j:
• Finally, it computes for each factor the average of
these deviation scores, and selects the contextual
factors with the largest average scores
ˆwuicj
= fcj
ˆruicj
− ˆrui
Normalized freq. of cj
Rating prediction when cj holds
Predicted context-free rating
15. ENTER 2016 Research Track Slide Number 16
Illustrative Example
• Let ȓAliceSkiingSunny = 5, ȓAliceSkiing = 3.5 & fSunny = 0.2. Then, the
impact of the “Sunny” condition is:
– ŵAliceSkiingSunny = 0.2 · |5 - 3.5| = 0.3
• Let ŵAliceSkiingCloudy= 0.2, ŵAliceSkiingRainy= 0.3 &
ŵAliceSkiingSnowy= 0.1, the impacts of the other weather
conditions. Then, the overall impact of the
“Weather” factor is:
– (0.3 + 0.2 + 0.3 + 0.1) ÷ 5 = 0.18
17. ENTER 2016 Research Track Slide Number 18
Datasets
Dataset STS TripAdvisor
Rating scale 1-5 1-5
Ratings 2,534 4,147
Users 325 3,916
Items 249 569
Contextual factors 14 3
Contextual conditions 57 31
Avg. # of factors known for
each rating
1.49 3
User attributes 7 2
Item attributes 1 12
In STS when a user rates
a POI she commonly
specifies at most 4 out
of the 14 factors!
In STS when a user rates
a POI she commonly
specifies at most 4 out
of the 14 factors!
18. ENTER 2016 Research Track Slide Number 19
Evaluation Procedure: Overview
• Repeated random sampling (20 times):
– Randomly partition the ratings into 3 subsets
– For each user-item pair (u, i) in the candidate set, compute the N most
relevant contextual factors and transfer the corresponding rating and
context information ruic in the candidate set to the training set as ruic’ with
c’ c containing the conditions for these factors, if any⊆
– Measure prediction accuracy (MAE) and ranking quality (Precision) on
testing set, after training the prediction model on extended training set
– Repeat
Training set (25%) Candidate set (50%) Testing set (25%)
19. ENTER 2016 Research Track Slide Number 20
user-item pair
Evaluation Procedure: Example
(Alice, Skiing)
Season and Weather
rAlice Skiing Winter, Sunny, Warm, Morning = 5
rAlice Skiing Winter, Sunny = 5
top two contextual factors
rating in candidate set
rating transferred to training set
+
+
=
20. ENTER 2016 Research Track Slide Number 21
Baseline Methods for Evaluation
• Mutual Information: given a user-item pair (u,i),
computes the relevance for a contextual factor Cj
as the mutual information between ratings for
items belonging to i’s category (Baltrunas et al., 2012)
• Freeman-Halton Test: calculates the relevance of
Cj using the Freeman-Halton test (Odić et al., 2013)
• mRMR: ranks each Cj according to its relevance to
the rating variable and redundancy to other
contextual factors (Peng et al., 2005)
21. ENTER 2016 Research Track Slide Number 22
Evaluation Results:
Prediction Accuracy
22. ENTER 2016 Research Track Slide Number 23
Evaluation Results:
Ranking Quality
23. ENTER 2016 Research Track Slide Number 24
Evaluation Results:
# of Acquired Conditions
25. ENTER 2016 Research Track Slide Number 26
Conclusions
• Using Largest Deviation, we know that we can ask
only the contextual factors C1, C2 and C3 when we
ask user u to rate item i