Slides from the paper Ladislav Peska, Peter Vojtas:Using Implicit Preference Relations to Improve Content-based Recommendations. on EC_WEB 2015 Conference (DEXA event), Valencia, Spain

1. Using Implicit Preference
Relations to Improve Content
Based Recommending
Ladislav Peška and Peter Vojtáš
Department of Software Engineering,
Charles University in Prague,
Czech Republic

2. Peska, Vojtas. Using IPR to Improve Content-
Based Recommending
2
Recommender Systems
 Propose relevant items to the right persons at the right
time
 Machine learning application
 Expose otherwise hard to find, uknown items
 Complementary to the catalogues, search engines etc.
 „Win-win strategy“
EC-WEB 2015, Valencia
User Feedback
rating, clickstream,
time on page, buys…
User, Object Profiles
Object attributes
(Context)
Time, location,
Possible choices…
RECOMMENDER
SYSTEM
Top-K Recommended objects

3. Peska, Vojtas. Using IPR to Improve Content-
Based Recommending
3
Recommender Systems
 User feedback
 Explicit feedback (rating)
 Implicit feedback (user behavior)
 Dwell time, clickstream, scrolling, mouse moves etc.
 Often used as a proxy to the user rating
 Recommending algorithms
 Collaborative filtering
 (Users A and B were similar so far, the should like similar things in the future too)
 Cold start problem
 Content-based filtering
 (User A should like similar items to the ones he liked so far)
 Overspecialization, lack of diversity, obvious recommendations…
EC-WEB 2015, Valencia

4. Peska, Vojtas. Using IPR to Improve Content-
Based Recommending
4
Challenge
 Recommending for small e-commerce websites
 Tens of similar vendors, user can choose whichever she likes
 (Almost) no explicit feedback
(No incentives for users)
 Few visited pages
(Often usage of external search engines & landing on object details)
 Low user loyalty
(New vs. Returning visitors ratio 80:20)
 Not enough data for collaborative filtering
 Focus on Implicit Feedback & Content-based recommendations
 Gather as much as possible user feedback; the sooner the better
 Gather external content to improve CB recommendations (other papers)
EC-WEB 2015, Valencia

5. Peska, Vojtas. Using IPR to Improve Content-
Based Recommending
5
User Feedback
 Explicit feedback (provided via website GUI)
 Rating an object via Likert Scale
 Comparing objects explicitly is not so common
 Implicit feedback (Virtually any JS event could be used)
 Actions related to evaluation of a single object
 Dwell time on the object detail page
 Number of page views
 Scrolling, mouse events
 Select / copy text, printing, purchase process etc.
 Actions related to evaluation of a list of objects
 Analyze user behavior on the category pages,
search results etc.
 Search related actions etc.
EC-WEB 2015, Valencia
A Bor
Results
Selected object IDs:
1,4
Ignored object IDs:
2,3,5,6,7,8

6. Peska, Vojtas. Using IPR to Improve Content-
Based Recommending
6
Our Working Hypothesis
 Users are often evaluating lists of objects
 Search results, category pages, recommended items etc.
 If user selects some objects from the list, we take it as an
evidence of his/her positive preference.
 User prefers selected object(s) more, than other displayed &
ignored objects
 We can form preference relations:
IPRrel (selected obj. > ignored obj.)
 We can extend such relations along the content-based
similarity of objects
 Some objects could be ignored, because user was not
aware of them, not becouse he/she did not like them
 E.g. they were displayed below the visible area
EC-WEB 2015, Valencia
>
>

7. Peska, Vojtas. Using IPR to Improve Content-
Based Recommending
7
Outline of Our Approach
EC-WEB 2015, Valencia

8. Peska, Vojtas. Using IPR to Improve Content-
Based Recommending
8
Collecting User Behavior
 IPIget component for collecting user behavior
 Browser visible area size
 List of all objects and its positions on the page
 Listener on Scrolling events
 Compute visible time for each displayed object, use it as a proxy to
the level of user evaluation
 Some more refined approaches are possible (e.g. registering mouse moves or
visual focus for different quadrants)
 Listener on Clicking events (which object(s) were selected by the user)
IPIget component download: http://ksi.mff.cuni.cz/~peska/ipiget.zip
EC-WEB 2015, Valencia

9. Peska, Vojtas. Using IPR to Improve Content-
Based Recommending
9
Collecting User Behavior – Example
EC-WEB 2015, Valencia

10. Peska, Vojtas. Using IPR to Improve Content-
Based Recommending
10
Extending IPR Relations
 IPR(Ox,Oy,intx,y)
EC-WEB 2015, Valencia

11. Peska, Vojtas. Using IPR to Improve Content-
Based Recommending
11
Using IPR to Reranking List of
Objects
EC-WEB 2015, Valencia

12. Peska, Vojtas. Using IPR to Improve Content-
Based Recommending
12
Using IPR to Reranking List of
Objects - Algorithm
EC-WEB 2015, Valencia

13. Peska, Vojtas. Using IPR to Improve Content-
Based Recommending
13
Using IPR to Reranking List of
Objects – Conflict Strategies
IPR(O4>O2): O4 is better than O2
 Forward:
 Move O4 just before O2
 Do not miss relevant objects
 Backward:
 Move O2 just after O4
 Do not show irrelevant objects
 Swap:
 Change positions of O4 and O2
 Keep objects well separated
EC-WEB 2015, Valencia
O1
O2
O3
O4
O5
O6
+ IPR(O4,O2,int)
O1
O4
O2
O3
O5
O6
O1
O3
O4
O2
O5
O6
O1
O4
O3
O2
O5
O6
Forward Backward Swap

14. Peska, Vojtas. Using IPR to Improve Content-
Based Recommending
14
Our Approach - Example
EC-WEB 2015, Valencia

15. Peska, Vojtas. Using IPR to Improve Content-
Based Recommending
15
Experiments
 Off-line experiments on Czech secondhand bookshop dataset
 1760 users, train set (2/3 of user data), test set (1/3)
 Recommender systems tries to predict visited objects
 Vector Space Model (VSM) with TF-IDF & Cosine similarity
 SimCat (recommending similar categories based on Collaborative Filtering)
 Stochastic Gradient Descent Matrix Factorization (SGD MF)
 nDCG and Presence@top-k metrics
EC-WEB 2015, Valencia
Method nDCG p@5 p@10 p@50
VSM + best IPR-rerank (sim:0.5, int:0.1, swap) 0.475 13.6% 15.7% 20.7%
VSM 0.464 13.2% 15.1% 19.6%
Best IPR-rank (sim:0.5, int:0.1, swap) 0.247 7.1% 7.7% 8.5%
SimCat + best IPR-rerank (sim:0.01, int:0.1, forward) 0.219 4.7% 6.3% 10.0%
SimCat 0.136 0.9% 1.5% 5.4%
SGD MF (500 lat. factors, max 500 iterations) 0.126 0.89% 1.2% 3.3%
Random recommendations 0.085 0.09% 0.14% 0.27%
MinSimilarity threshold, VSM
0.2 0.3 0.5 0.8
0.465 0.470 0.473 0.472
Conflict resolving, VSM
Forward Backward Swap
0.465 0.460 0.466

16. Conclusions, Future Work
 Implicit feedback could be more than just a substitution for user rating
 Collecting feedback on list of objects could give us insight about user decision
proces
 We used user behavior on list of objects to create Implicit
Preference Relations (IPR) between selected and ignored objects
 IPR can be extended along the object similarity axis
 We shown algorithm to update linear list of objects with IPRs
 IPR re-ranked recommendations outperformed original ones in an off-line
experiment
 Open Problems, Challenges
 How much was object really evaluated by the user? (Going beyond visibility)
 Which object features makes it desirable for the user? (Tailored object similarities)
 On-line deployment
EC-WEB 2015, Valencia Peska, Vojtas. Using IPR to Improve Content-
Based Recommending
16

17. EC-WEB 2015, Valencia Peska, Vojtas. Using IPR to Improve Content-
Based Recommending
17
Thank you!
Questions, comments?

18. Peska, Vojtas. Using IPR to Improve Content-
Based Recommending
18
Recommending in Czech
Second-hand Bookshop
 Mostly single item in stock
 Few content-based attributes (low information value)
- Title, author, price, category, textual description
- Hard to define informative attributes
- Title (and author name) in Czech
- No common book identifier
(ISBN mostly inapplicable)
 No explicit feedback
 Page-view, time on page, buys…
 Users identified through cookies
 Approx. 9500 active books
 50-100 visitors / day
 2-4 purchases
EC-WEB 2015, Valencia
RECOMMENDED
OBJECTS
CATEGORIES
Attributes
search
CATALOGUE