1. Context-aware Collaborative Filtering for Learning Interests of Mobile Users
George Lee
User Context-based Service Control Group
Network Laboratories
NTT DoCoMo R&D
4/16/2017

2. Overview
The Problem: Mobile users cannot easily get desired information
Proposed solution: automatic, personalized, context-aware event notification approach
Learns user interests
Recommends new information using collaborative filtering
Evaluation
The Problem: mobile users don’t know where to look to find information they need.
Related work: event notification, pub/sub, text retrieval, web portals, search engines, collaborative filtering
Research Challenges:
First part:
an efficient context-aware architecture for classifying and delivering events to mobile users based on detailed subscriptions,
An efficient, scalable matching algorithm for this system
Second part:
Learning context aware user subscriptions automatically, and unobtrusively
Key contributions:
1. An efficient context-aware architecture
2. An efficient matching system for complex event models
3. A context-aware collaborative filtering system to learn user preferences
My approach:
First, the architecture. Key points are,
it’s a pub/sub system
It automatically classifies events
It allows matching both new events and old events
It is context aware
The topic learner:
Learn user preferences
Recommends topics using CF, which is essential if there are many topics and contexts
My CF approach can take advantage of preference correlations between contexts
4/16/2017

3. Mobile users can’t easily get relevant information
Relevant information is:
Appropriate for their context
Personalized based on individual interests
Current and up-to-date
Static menu is inadequate
Too many choices
Difficult to navigate
Not personalized or context-aware
Drawbacks of information retrieval
Requires queries
Not good for new or changing information
Rather than thinking about the problem as helping users navigate a service list or menu, think of it as automatically supplying them with information proactively.
4/16/2017

4. Automatic, personalized, context-aware event notification
MIT News
CSAIL News
The Tech
Boston Dining News
Italian Restaurants …
Central Sq.
Sports
Red Sox Scores
Mobile Handset
Context:
Going to lab
CSAIL News:
Talk at 3pm G825
Central Sq. Dining:
New café opening
Red Sox vs. Yankees:
4-3 (6th inning) 1.
Learning
user
interests 2.
Efficiently
matching events
to users 3.
Determining
user context
Explain the terminology: events
There are three main features of this system
First point: this is personalized. The system knows that I don’t have a car and prefer taking the train, so it displays train schedules. Similarly, the sports, restaurant, and theme park information are also personalized.
Second point: it’s automatic. When an event occurs, the user is automatically notified. If the sports score changes, or if the wait time for the roller coaster changes, the information is automatically updated.
Third, it’s context-aware. If the user’s context changes, the information displayed will change.
This system uses the context, situation, task ontology, etc. of the SNS, but adds:
the capability for personalized information
automatic notification.
the ability to retrieve information from many sources
Automatic
Personalized
Context-aware
4/16/2017

5. Matching events and learning user interests
Matching Engine
decides which users match an event based on event descriptions and user interests
User Agent
automatically learns user
interests for the current
context based on user input
Event
User
Event
description
User
interests
User
input
Matching
Engine
User
Agent
Event
User
User
Event
4/16/2017

6. A user agent for learning user interests
automatically learns user
interests for the current
context based on user input
Context
Event
User
Event
description
User
interests
User
input
Matching
Engine
User
Agent
Event
User
Challenges:
Implicitly learning user interests
Recommending info in new contexts
User
Event
4/16/2017

7. Learning and automatically updating user interests
Automatically recommends new topics based on previous topic ratings
Context=
(location, task)
Implicitly learns user ratings for (context, topic) based on topics users select
User Agent
Mobile
Handset
User
Interests
Selected
Topics
Topic
Recommender
Topic Rating
Learner
Matching
Engine
Information = events, information organized by topic
Users choose topics for more info
Implicit rating
Context = location, task
location and task aware info
where this info comes from
other elements of context
topic rating learner
implicit rating. Define “rated”
alternative approaches
Recommendation
many topics, many contexts
outputs a list of user interests = topics
matching engine delivers events that match these interests
Event List
Topic 1
Topic 2
Topic 3
Matched events
4/16/2017

8. Making topic recommendations
Recommendations needed for new topics and contexts
Possible approaches:
Popularity: not personalized
Rating History: recommendations based on previous topic ratings
Collaborative Filtering (CF): recommendations based on interests of users with similar interests
4/16/2017

9. Context-aware Collaborative Filtering
Is User X interested in
“MIT News” for
context “Go to lab”?
Item = (context, topic)
To calculate a recommendation for item (C, T):
Find users who have rated (C, T)
Find users with similar interests
Decide whether to recommend (C, T) based on ratings of similar users
User A
Yes
User B
User X
User C
User D
4/16/2017

10. Collaborative filtering example (standard approach)
Similarity = 0.5 A B C X Y Z
User 2 A B C X Y Z
User 1
Similarity = 0.5 A B C X Y Z
User 3
4/16/2017

11. Enhanced approach: consider item similarity
Item = (context, topic)
Give more weight to ratings of similar items
Items with same context or same topic are more similar
Context similarity
Location (TGN)
Task
Topic similarity (ODP)
throw in a picture illustrating similarity between contexts, topics, etc.
4/16/2017

12. Collaborative filtering example (enhanced approach)
Item C: similarity = 1
Item Z: similarity = 0 A B C X Y Z
User 2 A B C X Y Z
User 1
Item C: similarity = 0 A B C X Y Z
Item Z: similarity = 1
User 3
4/16/2017

13. Recommender Evaluation
Evaluate ability of Enhanced CF to provide relevant information in a new context
User Interface: app on mobile handset
16 test subjects
8 for data collection
8 for evaluation
50 topics based on i-mode services
2 contexts
Going to see a movie in Tokyo
Going to Tokyo Disneyland
10 topics per recommender
Interleave topics from two recommenders and observe which topics users selected
vs. Random
vs. Rating History
vs. Standard CF
Recommender A
Recommender B
Mobile
Handset
Event List
Topic 1A
Topic 1B
Topic 2A
Topic 2B
Topic 3A
Topic 3B
4/16/2017

14. Effective: Enhanced CF can recommend relevant topics in new contexts
Compared to other approaches, enhanced CF topics selected
413% more than Random topics
49.7% more than Rating History topics
24.8% more than Regular CF topics
More studies needed to increase confidence
4/16/2017

15. Conclusion I proposed an event notification system for mobile users
Automatic
Personalized
Context-aware
Enhanced context-aware collaborative filtering topic recommender
Future work
Learning and recommendation algorithms
Incorporating confidence metrics
Context models
User studies
sparse topic & context spaces
how to deal with absolutely new topics & contexts
4/16/2017