We present our work in progress towards employing complex user feedback and its context in recommender systems. Our work is generally focused on small or medium-sized e-commerce portals. Due to the nature of such enterprises, explicit feedback is unavailable, but implicit feedback can be collected in both large amount and rich variety. However, some perceived values of implicit feedback may depend on the context of the page or user’s device (further denoted as presentation context). In this paper, we present an extended model of presentation context, propose methods integrating it into the set of implicit feedback features and evaluate these on the dataset of real e-commerce users. The evaluation corroborated the importance of leveraging presentation context in recommender systems.

1. Towards Complex User Feedback
and Presentation Context in
Recommender Systems
Peter Vojtas and Ladislav Peška
Department of Software Engineering,
Charles University in Prague,
Czech Republic

2. Peska, Vojtas: Towards Complex User
Feedback and Presentation Context in
Recommender Systems
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“
PPI 2017, Stuttgart, Germany
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: Towards Complex User
Feedback and Presentation Context in
Recommender Systems
3
Recommender Systems
 User feedback
 Explicit feedback (user rating)
 Implicit feedback (user behavior)
 User visited/bought object
 Usually binary feature
 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…
PPI 2017, Stuttgart, Germany

4. Peska, Vojtas: Towards Complex User
Feedback and Presentation Context in
Recommender Systems
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,
continuous cold-start problem
 Focus on Implicit Feedback & Content-based recommendations
 Obtain precise implicit user feedback in enough quantity, as early as possible
 Gather external content to improve CB recommendations (other papers)
PPI 2017, Stuttgart, Germany

5. User Feedback
Explicit feedback
 Provided via website GUI
 Rating an object via Likert Scale
 Missing in small E-Commerces
Implicit feedback
 Often binary in the literature
 User visited object
 User bought object
PPI 2017, Stuttgart, Germany Peska, Vojtas: Towards Complex User
Feedback and Presentation Context in
Recommender Systems
5

6. User Feedback
Explicit feedback
 Provided via website GUI
 Rating an object via Likert Scale
 Comparing objects explicitly is
not so common
 Missing in small E-Commerces
Implicit feedback
 Often binary in the literature
 User visited object
 User bought object
 Virtually any event triggered by
user could be feedback
 Get better picture about user
engagement / preference
PPI 2017, Stuttgart, Germany Peska, Vojtas: Towards Complex User
Feedback and Presentation Context in
Recommender Systems
6

7. User Feedback
Explicit feedback
 Provided via website GUI
 Rating an object via Likert Scale
 Comparing objects explicitly is
not so common
 Missing in small E-Commerces
Implicit feedback
 Virtually any event could be
used as feedback
 Tracked via JavaScript
 Dwell time
 Number of page views, Scrolling,
mouse events, copy text, printing
 Purchase process etc.
 Purchases represents fully
positive feedback
PPI 2017, Stuttgart, Germany Peska, Vojtas: Towards Complex User
Feedback and Presentation Context in
Recommender Systems
7

8. User Feedback
Software: Peska, IPIget: The Component for Collecting Implicit User Preference Indicators
PPI 2017, Stuttgart, Germany Peska, Vojtas: Towards Complex User
Feedback and Presentation Context in
Recommender Systems
8

9. User Feedback – Past Resarch
 Combine multiple implicit feedback features to estimate user rating
 Standard CB / CF recommender systems can be used afterwards
 Improvements over the usage of simple implicit feedback
Peska, Vojtas: How to Interpret Implicit User Feedback?
Peska, Eckhardt, Vojtas: Preferential Interpretation of Fuzzy Sets in E-shop Recommendation with Real Data Experiments
PPI 2017, Stuttgart, Germany Peska, Vojtas: Towards Complex User
Feedback and Presentation Context in
Recommender Systems
9
Dwell time: 10s
Scrolling: 100px
Mouse movement:250px
Dwell time: 100s
Scrolling: 200px
Mouse movement:450px
Rating: 0.2
Rating: 0.8

10. User Feedback – Past Resarch
 Combine multiple implicit feedback features to estimate user rating
 Standard CB / CF recommender systems can be used afterwards
 Improvements over the usage of simple implicit feedback
Is that all we can do?
PPI 2017, Stuttgart, Germany Peska, Vojtas: Towards Complex User
Feedback and Presentation Context in
Recommender Systems
10
Dwell time: 10s
Scrolling: 100px
Mouse movement:250px
Dwell time: 100s
Scrolling: 200px
Mouse movement:450px
Rating: 0.2
Rating: 0.8

11. Context of User Feedback
 Combine multiple implicit feedback features to estimate user rating
 Is that all we can do?
 Pages may substantially vary in length, amount of content etc.
 This could affect perceived implicit feedback features
 Leveraging context could be important
PPI 2017, Stuttgart, Germany Peska, Vojtas: Towards Complex User
Feedback and Presentation Context in
Recommender Systems
11
Dwell time: 10s
Scrolling: 100px
Mouse movement:250px
Dwell time: 100s
Scrolling: 200px
Mouse movement:450px
Rating: 0.2
Rating: 0.8

12. Peska, Vojtas: Towards Complex User
Feedback and Presentation Context in
Recommender Systems
12
Context of User Feedback
PPI 2017, Stuttgart, Germany
A B
 Context of the user
 Location, Mood, Seasonality...
 Can affect user preference
 Out of scope of this paper
 Context of device and page
 Page and browser dimensions
 Page complexity (amount of text, links, images,...)
 Device type
 Datetime
 Can affect percieved values of the user feedback

13. Outline of Our Approach
Traditional recommender
 User rates a sample of objects
 Preference learning computes
expected ratings of all objects
 Top-k best rated objects are
recommended
Our approach
PPI 2017, Stuttgart, Germany Peska, Vojtas: Towards Complex User
Feedback and Presentation Context in
Recommender Systems
13
𝑟𝑢,𝑜: 𝑜 ∈ 𝑺 ⊂ 𝑶; 𝑟𝑢,𝑜 ∈ [0,1]
𝑅 𝑢→ 𝑟𝑢,𝑜′ ∶ 𝑜′
∈ 𝑶
𝑅 𝑢 = {𝑜1, … , 𝑜 𝑘
𝐹𝑢,𝑜 = [𝑓1, … , 𝑓𝑖
𝑅 𝑢→ 𝑟𝑢,𝑜′ ∶ 𝑜′∈ 𝑶
𝐶 𝑢,𝑜 = [𝑐1, … , 𝑐𝑗
𝐹𝑢,𝑜, 𝐶 𝑢,𝑜 → 𝑟𝑢,𝑜: 𝑜 ∈ 𝑺
 Several imlicit feedback and contextual
features are collected:
 Learn estimated rating 𝑟𝑢,𝑜 for visited
objects based on feedback and context

 „The more the better” heuristics (STD, CDF)
 Machine learning approach (J48)
 Incorporate context
 As further feedback features (FB+C)
 As baseline predictors (AVGBP, CBP)
 Learn rating on all objects as in traditional
recommenders

14. Peska, Vojtas: Towards Complex User
Feedback and Presentation Context in
Recommender Systems
14
Collecting User Behavior
 IPIget component for collecting user behavior
IPIget component download: http://ksi.mff.cuni.cz/~peska/ipiget.zip
PPI 2017, Stuttgart, Germany
Contextual features
𝒄 𝟏 Number of links
𝒄 𝟐 Number of images
𝒄 𝟑 Text size
𝒄 𝟒 Page dimensions
𝒄 𝟓 Visible area ratio
𝒄 𝟔 Hand-held device
Implicit Feedback Features
𝒇 𝟏 View Count
𝒇 𝟐 Dwell Time
𝒇 𝟑,𝟒 Mouse Distance and Time
𝒇 𝟓,𝟔 Scrolled Distance and Time
𝒇 𝟕 Clicks count
𝒇 𝟖 Hit bottom of the page
𝒓 Purchase

15. Peska, Vojtas: Towards Complex User
Feedback and Presentation Context in
Recommender Systems
15
Estimated Rating from Implicit
Feedback
PPI 2017, Stuttgart, Germany
 „The more the better” heuristics
 Various feedback features are not comparable in general
 Dwell time (sec) vs. Distance travelled by mouse (pixels)
 Transform feedback features on comparable scale and average
 Use standardization (STD) of feedback features (≈ N(0,1))
 Use cummulative distribution (CDF) of each feedback feature
=>

16. Peska, Vojtas: Towards Complex User
Feedback and Presentation Context in
Recommender Systems
16
Estimated Rating from Implicit
Feedback
PPI 2017, Stuttgart, Germany
 Machine learning approach
 J48 decision tree
 Purchases are golden standard
 The only feedback which is a true indicator of positive preference
 Predict purchases based on other feedback features
 Use probability of purchase as estimated rating 𝑟𝑢,𝑜

17. Peska, Vojtas: Towards Complex User
Feedback and Presentation Context in
Recommender Systems
17
Employ Context in Rating
Estimation
PPI 2017, Stuttgart, Germany
 Use context in the same way as feedback (FB+C)
 Leave the decision about usage of context on the underlined model
 Plausible strategy for e.g. decision trees or rule mining learning
approaches
 Use context as a baseline predictor of feedback
 Calculate estimated value of feedback feature for particular context
value 𝑓𝑖(𝑐𝑗)
 Substract the estimation from the actual value 𝑓𝑖.𝑢.𝑜
𝑏𝑝
= 𝑓𝑖,𝑢,𝑜 − 𝑓𝑖 𝑐𝑗
 Use feedback with baseline estimators instead of the original one
 Either employ average baseline predictor over all context features
(AVGBP)
 Or use carthesian product of feedback features and baseline
predictors based on each context feature (CBP)

18. Peska, Vojtas: Towards Complex User
Feedback and Presentation Context in
Recommender Systems
18
Preference Learning and
Recommendations
PPI 2017, Stuttgart, Germany
 Collaborative filtering not applicable
 Continuous cold-start problem
 Use combination of content-based and non-personalized
 VSM content-based recommendation
 Vector of object features (TF-IDF)
 User is represented as weighted sum of visited object’s features
 Resulting score is a cosine similarity of user and object vectors
 Most popular non-personalized algorithm
 Based on estimated ratings 𝑟𝑢,𝑜
 Final score 𝑟𝑢,𝑜′ is a multiplication of VSM score and most
popular score

19. Peska, Vojtas: Towards Complex User
Feedback and Presentation Context in
Recommender Systems
19
Evaluation
PPI 2017, Stuttgart, Germany
 Czech travel agency dataset
 3 variants of rating estimation (STD, CDF, J48)
 3 variants of context incorporation (FB+C, AVGBP, CBP)
 2 baselines (use raw feedback, use binary visits)
 Leave-one-out on purchased objects
 Ranking prediction
 nDCG, recall@top-10

20. Peska, Vojtas: Towards Complex User
Feedback and Presentation Context in
Recommender Systems
20
Results
 Results of nDCG, (*) = signifficant improvement of the best method
 J48 decision tree with both feedback and context on its input performs the
best
 Using „the more the better“ heuristics (CDF) with properly processed
feedback (AVGBP, CBP) also performs quite well
PPI 2017, Stuttgart, Germany
Processing
method
Feedback and Context composition
Binary FB FB+C AVGBP CBP
STD + popVSM 0.255* 0.174* 0.197* 0.161* 0.158*
CDF + popVSM 0.255* 0.257* 0.253* 0.258* 0.257
J48 + popVSM 0.255* 0.256* 0.274 0.240* 0.247*
J48 + objects
popularity
0.180** 0.205* 0.211* 0.168* 0.186*
J48 + VSM 0.222* 0.224* 0.233 0.225* 0.224*

21. Conclusions, Future Work
Key outcomes
 Implicit feedback could be more than just a binary variable
 Observed feedback should be considered with respect to the context of
page and device
 Doing so could improve the quality of the recommended objects
Future work, Open Problems
 Better models of context employment and purchase prediction methods
 Further evaluation scenarios
 Recommending on the beginning of a new session
 More refined feedback?
 E.g. feedback on object’s attributes?
 On-line deployment and evaluation
PPI 2017, Stuttgart, Germany Peska, Vojtas: Towards Complex User
Feedback and Presentation Context in
Recommender Systems
21

22. PPI 2017, Stuttgart, Germany Peska, Vojtas: Towards Complex User
Feedback and Presentation Context in
Recommender Systems
22
Thank you!
Questions, comments?
Supplementary materials: http://bit.ly/2g79VVO
Slides: