Upload
forgetit-project
View
143
Download
3
Tags:
Embed Size (px)
Citation preview
Concise Preservation by combining Managed
Forgetting and Contextualized Remembering
Nattiya Kanhabua
L3S Research Center
WP 3 Presentation
Managed Forgetting
ForgetIT 1st Review Meeting, April 29-30, 2014
Kaiserslautern, Germany
WP3 Objectives
• Conceptual model for managed forgetting Foundations of human-brain inspired managed forgetting
• Development of managed forgetting methods Information value assessment
Set of methods for Preserve-or-Forget
Policy-driven approach to managed forgetting (Y2)
Focus of Year 1
• Conceptual model for managed forgetting
• Design and implement the core managed forgetting process
• Exploratory research of information value assessment
ForgetIT Project GA600826, 1st Review Meeting, Kaiserslautern, April 2014
Objectives of WP3 and Year 1 Focus
ForgetIT Project GA600826, 1st Review Meeting, Kaiserslautern, April 2014
Role in Preserve-or-Forget Architecture
Research questions and first ideas for complementing human memory
(co-worked with WP2, D3.1) • Episodic memory: reconstruct lifetime memories and support reminiscence
• Working memory: better focus in current information use
Information value assessment (co-worked with WP9, D3.2)
• Data model and a computation method based on Semantic Web technologies
• Integration to PIMO semantic desktop and Preserve-or-Forget middleware
Exploratory studies (D3.2)
• Analyzing collective memory of public events in Wikipedia
• Analyzing high-impact features for content retention in the Social Web
• Feature selection for efficiency and scalability
ForgetIT Project GA600826, 1st Review Meeting, Kaiserslautern, April 2014
Achievements in Year 1
Goal: understand how to complement human memory processes
Focus on two types of memories:
• Episodic memory: support reminiscence of long-term autobiographical events
• Working memory: better focus in current information use, e.g. de-cluttering
personal information spaces
Two information values: memory buoyancy, and preservation value
ForgetIT Project GA600826, 1st Review Meeting, Kaiserslautern, April 2014
Complementing Human Memory: Our First Ideas
Memory buoyancy
• Information objects sinking down with decreasing importance, usage, etc.
Preservation value
• Used to decide which information object will be preserved or archived
ForgetIT Project GA600826, 1st Review Meeting, Kaiserslautern, April 2014
Information Value Assessment
Memory Buoyancy Preservation Value
Short-/Mid-term current interests
E.g. meeting or travel documents
Long-term need for future use
E.g. important life events
Subjective metrics
+ usage logs (views, edits, modifies)
+ time, e.g., aging or recency
+ social context, external influences
Objective metrics
+ diversity, coverage, quality
Rapidly forget details -> “less redundancy”
Reconstruct from similar events, context
Rely on common patterns -> “false memory”
Our first ideas:
• Store details differing among similar event types forgotten in human memory
• Event-centric organization of digital items can play an important role
ForgetIT Project GA600826, 1st Review Meeting, Kaiserslautern, April 2014
Forgetting in Episodic Memory
Memory bumps or peaks in the forgetting curve
Reminded or triggered the original memory by:
• A physical object (e.g. a printed photo)
• A digital memory system
• Different subsequent events
Our ideas:
• Propagate increased interest in an event to related events
• Consider common things, e.g., same entities, or similar event types
• Increase relevance level or use of memory buoyancy
ForgetIT Project GA600826, 1st Review Meeting, Kaiserslautern, April 2014
Triggering of Memories
ForgetIT Project GA600826, 1st Review Meeting, Kaiserslautern, April 2014
Analyzing Collective Memory in Wikipedia
Identify catalysts for reviving memories
Analyze re-visiting behaviors
• Page views of a large set of events
• Time series analysis
11 Wikipedia categories
• Number of triggering events
• Number of events possibly triggered
Temporal and spatial distributions
• Strong focus on more recent events
• Better coverage with increasing popularity
• Most frequent locations depending on event types
ForgetIT Project GA600826, 1st Review Meeting, Kaiserslautern, April 2014
Temporal and Spatial Distributions
ForgetIT Project GA600826, 1st Review Meeting, Kaiserslautern, April 2014
Our Approach and Results
Remembering score as a function (e.g., detecting co-peaks in views) of revisiting behavior
Correlate remembering scores vs. time and location similarities
Hurricane Sandy Findings:
• Hurricane Sandy triggers 1991 Perfect Storm
initially formed around Canada area, which is
high impact (most destructive and costly) ones
• 2011 Christchurch earthquake triggers recent
events in the same region, i.e., 2010 Canterbury
earthquake
ForgetIT Project GA600826, 1st Review Meeting, Kaiserslautern, April 2014
Our Approach and Results
Remembering score as a function (e.g., detecting co-peaks in views) of revisiting behavior
Correlate remembering scores vs. time and location similarities
Hurricane Sandy 2011 Christchurch earthquake Findings:
• Hurricane Sandy triggers 1991 Perfect Storm
initially formed around Canada area, which is
high impact (most destructive and costly) ones
• 2011 Christchurch earthquake triggers recent
events in the same region, i.e., 2010 Canterbury
earthquake
ForgetIT Project GA600826, 1st Review Meeting, Kaiserslautern, April 2014
Memory Buoyancy: Simplified Computation
Me
mo
ry B
uo
ya
nc
y
Time
Compute: MB(D, t)
Time
Ac
ce
ss
Lo
gs
t1 t2
Proposed MB assessment framework:
• Initialize MB values of resources
using a time-decay forgetting function:
• Incrementally update MB using
Random Walk on resource graph:
ForgetIT Project GA600826, 1st Review Meeting, Kaiserslautern, April 2014
Memory Buoyancy Assessment
|'|)( )( ttt DecayRatermb
r
e2
Edfringe photo (2011)
Photos @ iPhone
e3
Folder @ computer
e1
Shortcut folder @ desktop
e4 e6
Photo @ ForgetIT Meeting (2013)
contains
contains
contains
hasSamePlace
hasSamePlace
e5 hasEntity
Whiskey photo (2012)
2
)(
1
)(
2
1)( 4
)(6
)()1( embemb
rmbt
Dasht
DashtDash
Averaged value over
two inlinked resources
Less propagation
account for two outlinks
hasSamePlace
e5
Whiskey Tour (2009)
hasSamePlace
Social Web apps gain popularity
Personal Web archives
Study: Identifying memorable content • 20 participants, 15 male and 5 female
• Rate (3,330) posts by relevance for future
ForgetIT Project GA600826, 1st Review Meeting, Kaiserslautern, April 2014
Content Retention in Social Web Applications
Year in Review: photo from the Internet
Machine learning techniques
• Support vector machine, Bayesian network, and decision tree (J48)
80 features from categories:
• Content types + meta data
• Social interactions
• Temporal
• Privacy
• Graph
Correlation-based feature selection (CFS) • Temporal: highest impact features
• Graph: low impact for memorable posts
ForgetIT Project GA600826, 1st Review Meeting, Kaiserslautern, April 2014
Learning to Classify Memorable Content
Classification results: • Baseline Features (CS): No. of likes, comments, and shares
• Baseline 69% (F-Measure)
• Top 9 features 79% (F-Measure)
ForgetIT Project GA600826, 1st Review Meeting, Kaiserslautern, April 2014
Classification Results
1. M. Georgescu, D. D. Pham, N. Kanhabua, S. Zerr, S. Siersdorfer and W. Nejdl, Temporal Summarization of
Event-Related Updates in Wikipedia (demo), Proceedings of the 22nd International World Wide Web Conference
(WWW'13), May, 2013.
2. M. Georgescu, N. Kanhabua, D. Krause, W. Nejdl and S. Siersdorfer, Extracting Event-Related Information from
Article Updates in Wikipedia, Proceedings of the 35th European conference on Advances in Information Retrieval
(ECIR'13), March, 2013.
3. N. Kanhabua and C. Niederée, Preservation and Forgetting: Friends or Foes?, In the First International
Workshop on Archiving Community Memories (in conjunction with iPRES'2013), September, 2013.
4. N. Kanhabua, C. Niederée and W. Siberski, Towards Concise Preservation by Managed Forgetting: Research
Issues and Case Study, Proceedings of the 10th International Conference on Preservation of Digital Objects
(iPRES'2013), September, 2013.
5. K. D. Naini and I.S. Altingovde, Exploiting Result Diversification Methods for Feature Selection in Learning to
Rank, Proceedings of the 36th European conference on Advances in Information Retrieval (ECIR'2014), April, 2014.
6. A. Ceroni and M. Fisichella, Towards an Entity-based Automatic Event Validation, Proceedings of the 36th
European conference on Advances in Information Retrieval (ECIR'2014), April, 2014.
7. T. N. Nguyen and N. Kanhabua, Leveraging Dynamic Query Subtopics for Time-aware Search Result
Diversification, Proceedings of the 36th European conference on Advances in Information Retrieval (ECIR'2014),
April, 2014.
8. K. D. Naini, R. Kawase, N. Kanhabua and C. Niederée, Characterizing High-impact Features for Content
Retention in Social Web Applications (poster), Proceedings of the 23rd International World Wide Web
Conference (WWW'2014), Seoul, Korea, April, 2014.
9. T. A. Tran, M. Georgescu, X. Zhu and N. Kanhabua, Ars longa, vita brevis: Analysing the Duration of Trending
Topics in Twitter Using Wikipedia (poster), (To appear) Proceedings of the ACM Web Science 2014 Conference
(WebSci'2014), Bloomington, USA, June, 2014.
ForgetIT Project GA600826, 1st Review Meeting, Kaiserslautern, April 2014
Publications
Thank you for your attention!