User-Oriented Context Suggestion
Yong Zheng, Bamshad Mobasher, Robin Burke
Center for Web Intelligence
DePaul University, Chicago
The 24th Conference on User Modeling, Adaptation and Personalization
Halifax, Canada, July 13-16, 2016
Question: Is it enough to have
appropriate/good
item recommendations?
Zoo Parks in San Diego, USA
3
• San Diego Zoo • San Diego Zoo Safari Park
Zoo Parks in San Diego, USA
4
Intro: Context Suggestion
Traditional Recommender Systems
6
• Task: Suggest a list of items to a user
• For example, recommend me a list of movies to watch
Traditional Rec
Context-aware Recommendation
7
• Task: Suggest a list of items to a user in specific contexts
• For example, recommend me some movies to watch with my
girlfriend at weekend in the cinema
Contextual RecTraditional Rec
Context Suggestion
8
• Task: Suggest a list of contexts to users/items
• For example: suggest me the time/location to watch a movie
Context Rec
Contextual RecTraditional Rec
What is Context?
9
• Our definition:
Contexts are those variables which may change when a same
activity is performed repeatedly – not only the time & location,
but also companion, occasions, user intent/purpose, etc
• Examples:
Watching a movie: time, location, companion, etc
Listening to a music: time, location, emotions, occasions, etc
Party or Restaurant: time, location, occasion, etc
Travels: time, location, weather, transportation condition, etc
Motivations
11
Context Suggestion: Motivations
• Motivation-1: Maximize user experience
User Experience (UX) refers to a person's emotions and
attitudes about using a particular product, system or
service.
12
Context Suggestion: Motivations
• Motivation-1: Maximize user experience
It is not enough to recommend good items only
Good item recommendations cannot guarantee the whole user experience!
13
Context Suggestion: Motivations
• Motivation-2: Contribute to Context Collection
Predefine contexts and suggest them to users
14
Context Suggestion: Motivations
• Motivation-3: Connect with Context-aware RecSys
User’s actions on context is a context-query to system
Applications
16
Context Suggestion: Applications
• There could be many potential applications:
17
Context Suggestion
• There could be many applications, we focus on two tasks
1).UI-Oriented Context Suggestion
Task: suggest contexts to <user, item>
Example: time & location for me to watch Life of Pi
Existing solutions: Multi-label classification/predictions
2). User-Oriented Context Suggestion
Task: suggest contexts to each user
Example: Google Music, Pandora, Youtube, etc
Solution: this paper in UMAP 2016
Challenges and Solutions
19
Challenge: Evaluations
1).UI-Oriented Context Suggestion
Task: suggest contexts to <user, item>
Example: time & location for me to watch Life of Pi
2). User-Oriented Context Suggestion
Task: suggest contexts to each user
Example: Google Music, Pandora, Youtube, etc
Same challenge: Evaluations!!
We do not have user’s preferences on contexts. No data!
20
Evaluation: Solutions
In this paper, we use a simulation-based approach.
User’s taste on context
conditions can be obtained
by the average rating on
context condition by users
across contextual ratings
over all rated items.
21
Algorithms: User-Oriented Context Suggestion
Solution 1). By Contextual Rating Deviations (CRDs)
CRD is used to tell how user’s rating is deviated in each context
condition. For example, CRD(u, weekend) = 0.5,
it tells that user u’s rating on items is usually higher by 0.5
if watching movies at weekend
CAMF_C:
CAMF_CU:
22
Algorithms: User-Oriented Context Suggestion
Solution 2).By UI-Oriented Context Suggestion
Color, Shape, Weight, Origin,
Taste, Price, Vitamin c
Predictions By UI-Oriented Context Suggestion
Converted User-Context Predictions
23
Algorithms: User-Oriented Context Suggestion
Solution 2).By UI-Oriented Context Suggestion
We choose two methods in UI-Oriented context suggestion
I). Multi-Label Classification (MLC)
We use LabelPowerset (LP) + RandomForest
II). Tensor Factorization (PITF)
Color, Shape, Weight, Origin,
Taste, Price, Vitamin c
Users × Items × Contexts Ratings
Results and Findings
We present the results based on the music data: 42 users, 139
items, 3938 ratings, 34 contexts to be suggested. We examine
top-5 suggestions by the 5-fold cross validation.
24
Simple Baseline By Context Rating Deviations By UI Context Suggestion
Results and Findings
We present the results based on the music data: 42 users, 139
items, 3938 ratings, 34 contexts to be suggested. We examine
top-5 suggestions by the 5-fold cross validation.
25
Simple Baseline By Context Rating Deviations By UI Context Suggestion
Conclusions and Future Work
• Conclusions
PITF is the best algorithm; MLC is ranked the 2nd
CRD-based approach works better than baseline
• Future Work
Collect appropriate data & Perform user studies
Try other contextual recommendation algorithms
• Acknowledgement
 Student Travel Grant by US NSF
26
User-Oriented Context Suggestion
Yong Zheng, Bamshad Mobasher, Robin Burke
Center for Web Intelligence
DePaul University, Chicago
The 24th Conference on User Modeling, Adaptation and Personalization
Halifax, Canada, July 13-16, 2016

[UMAP 2016] User-Oriented Context Suggestion

  • 1.
    User-Oriented Context Suggestion YongZheng, Bamshad Mobasher, Robin Burke Center for Web Intelligence DePaul University, Chicago The 24th Conference on User Modeling, Adaptation and Personalization Halifax, Canada, July 13-16, 2016
  • 2.
    Question: Is itenough to have appropriate/good item recommendations?
  • 3.
    Zoo Parks inSan Diego, USA 3 • San Diego Zoo • San Diego Zoo Safari Park
  • 4.
    Zoo Parks inSan Diego, USA 4
  • 5.
  • 6.
    Traditional Recommender Systems 6 •Task: Suggest a list of items to a user • For example, recommend me a list of movies to watch Traditional Rec
  • 7.
    Context-aware Recommendation 7 • Task:Suggest a list of items to a user in specific contexts • For example, recommend me some movies to watch with my girlfriend at weekend in the cinema Contextual RecTraditional Rec
  • 8.
    Context Suggestion 8 • Task:Suggest a list of contexts to users/items • For example: suggest me the time/location to watch a movie Context Rec Contextual RecTraditional Rec
  • 9.
    What is Context? 9 •Our definition: Contexts are those variables which may change when a same activity is performed repeatedly – not only the time & location, but also companion, occasions, user intent/purpose, etc • Examples: Watching a movie: time, location, companion, etc Listening to a music: time, location, emotions, occasions, etc Party or Restaurant: time, location, occasion, etc Travels: time, location, weather, transportation condition, etc
  • 10.
  • 11.
    11 Context Suggestion: Motivations •Motivation-1: Maximize user experience User Experience (UX) refers to a person's emotions and attitudes about using a particular product, system or service.
  • 12.
    12 Context Suggestion: Motivations •Motivation-1: Maximize user experience It is not enough to recommend good items only Good item recommendations cannot guarantee the whole user experience!
  • 13.
    13 Context Suggestion: Motivations •Motivation-2: Contribute to Context Collection Predefine contexts and suggest them to users
  • 14.
    14 Context Suggestion: Motivations •Motivation-3: Connect with Context-aware RecSys User’s actions on context is a context-query to system
  • 15.
  • 16.
    16 Context Suggestion: Applications •There could be many potential applications:
  • 17.
    17 Context Suggestion • Therecould be many applications, we focus on two tasks 1).UI-Oriented Context Suggestion Task: suggest contexts to <user, item> Example: time & location for me to watch Life of Pi Existing solutions: Multi-label classification/predictions 2). User-Oriented Context Suggestion Task: suggest contexts to each user Example: Google Music, Pandora, Youtube, etc Solution: this paper in UMAP 2016
  • 18.
  • 19.
    19 Challenge: Evaluations 1).UI-Oriented ContextSuggestion Task: suggest contexts to <user, item> Example: time & location for me to watch Life of Pi 2). User-Oriented Context Suggestion Task: suggest contexts to each user Example: Google Music, Pandora, Youtube, etc Same challenge: Evaluations!! We do not have user’s preferences on contexts. No data!
  • 20.
    20 Evaluation: Solutions In thispaper, we use a simulation-based approach. User’s taste on context conditions can be obtained by the average rating on context condition by users across contextual ratings over all rated items.
  • 21.
    21 Algorithms: User-Oriented ContextSuggestion Solution 1). By Contextual Rating Deviations (CRDs) CRD is used to tell how user’s rating is deviated in each context condition. For example, CRD(u, weekend) = 0.5, it tells that user u’s rating on items is usually higher by 0.5 if watching movies at weekend CAMF_C: CAMF_CU:
  • 22.
    22 Algorithms: User-Oriented ContextSuggestion Solution 2).By UI-Oriented Context Suggestion Color, Shape, Weight, Origin, Taste, Price, Vitamin c Predictions By UI-Oriented Context Suggestion Converted User-Context Predictions
  • 23.
    23 Algorithms: User-Oriented ContextSuggestion Solution 2).By UI-Oriented Context Suggestion We choose two methods in UI-Oriented context suggestion I). Multi-Label Classification (MLC) We use LabelPowerset (LP) + RandomForest II). Tensor Factorization (PITF) Color, Shape, Weight, Origin, Taste, Price, Vitamin c Users × Items × Contexts Ratings
  • 24.
    Results and Findings Wepresent the results based on the music data: 42 users, 139 items, 3938 ratings, 34 contexts to be suggested. We examine top-5 suggestions by the 5-fold cross validation. 24 Simple Baseline By Context Rating Deviations By UI Context Suggestion
  • 25.
    Results and Findings Wepresent the results based on the music data: 42 users, 139 items, 3938 ratings, 34 contexts to be suggested. We examine top-5 suggestions by the 5-fold cross validation. 25 Simple Baseline By Context Rating Deviations By UI Context Suggestion
  • 26.
    Conclusions and FutureWork • Conclusions PITF is the best algorithm; MLC is ranked the 2nd CRD-based approach works better than baseline • Future Work Collect appropriate data & Perform user studies Try other contextual recommendation algorithms • Acknowledgement  Student Travel Grant by US NSF 26
  • 27.
    User-Oriented Context Suggestion YongZheng, Bamshad Mobasher, Robin Burke Center for Web Intelligence DePaul University, Chicago The 24th Conference on User Modeling, Adaptation and Personalization Halifax, Canada, July 13-16, 2016