Reasoning about Weighted Semantic User Profiles through Collective Confidence Analysis: A Fuzzy Evaluation Nima Dokoohaki Mihhail Matskin Presented at 6 th Atlantic Web Intelligence Conference (AWIC09) September 2009

Motivation Introduction to Domain Problem statement Processing the weights Confidence Definition Modeling Confidence evaluation process Step-by-step Pictorial Evaluation Experiment Setting Experiment results Analysis Conclusions Agenda

Motivation

Introduction to Domain

Problem statement

Processing the weights

Confidence

Definition

Modeling

Confidence evaluation process

Step-by-step

Pictorial

Evaluation

Experiment Setting

Experiment results

Analysis

Conclusions

Profiled material include: Interest capturing known as a traditional approach in user profiling assigning extra weights for capturing trust, privacy and rank with respect to item category user has interest towards or with respect to each item user has visited. Motivation for assigning weights to profiled items Altering or (ultimately) improving system performance based on user gathered data Profiled Content= Personal data+ Weights

Profiled material include:

Interest capturing

known as a traditional approach in user profiling

assigning extra weights for capturing trust, privacy and rank with respect to item category user has interest towards or with respect to each item user has visited.

Motivation for assigning weights to profiled items

Altering or (ultimately) improving system performance based on user gathered data

RFID information tags RFID information tags SMARTMUSEUM Servers Museum Artifacts User with handheld Mobile device Museum Artifacts SMARTMUSEUM Depicted

Profile sample Example: <http://www.smartmuseum.eu/ns/context/weather#, visited, Saint_Jerome_Writing, atDate 20081210, 0.8, 0.6, 0.5> We interpret this as: “ With reference to Weather ( context ), at a certain date (20081210), user visited Saint Jerome Writing artwork and liked it very much ( rank value) and user trusts (the originality of) the work visited and has average privacy (consensus on the disclosure of the information to the public).” Saint Jerome Writing Artist Caravaggio Year c. 1607-1608 Type Oil on canvas Dimensions 117 cm × 157 cm (46 in × 62 in) Location St John's Co-Cathedral, Malta

Example:

<http://www.smartmuseum.eu/ns/context/weather#,

visited, Saint_Jerome_Writing, atDate 20081210,

0.8, 0.6, 0.5>

We interpret this as:

“ With reference to Weather ( context ), at a certain date (20081210), user visited Saint Jerome Writing artwork and liked it very much ( rank value) and user trusts (the originality of) the work visited and has average privacy (consensus on the disclosure of the information to the public).”

In order to improve the system performance we have to analyze weights assigned to profiled data. Problem Statement: Processing profiled Weights (Trust, Privacy, Rank) Values “ How to process/evaluate the weights from profiled data?” “ What factors we have to consider for a meaningful interpretation of the process?” Specific problem Introduced

In order to improve the system performance we have to analyze weights assigned to profiled data.

Problem Statement:

Processing profiled Weights

(Trust, Privacy, Rank) Values

“ How to process/evaluate the weights from profiled data?”

“ What factors we have to consider for a meaningful

interpretation of the process?”

Confidence is The state of being certain Certainty of an experience is affected by situation-dependent measures People may have confidence in Other people or forces beyond their control. Trust, belief and confidence are synonymous when used in the same context. Confidence

Confidence is

The state of being certain

Certainty of an experience is affected by situation-dependent measures

People may have confidence in

Other people or

forces beyond their control.

Trust, belief and confidence are synonymous when used in the same context.

Aligning / Adopting Confidence to profiling domain Usage Confidence The state of being certain … with respect to values given as input to mobile devices during museum visit. Certainty of visiting experience is affected by situation-dependent measures Weights (trust , privacy and ranks)

Usage Confidence

The state of being certain … with respect to values given as input to mobile devices during museum visit.

Certainty of visiting experience is affected by situation-dependent measures

Weights (trust , privacy

and ranks)

Individual Confidence Modeling: Calculating confidence for a specific user, with respect to item/items which user has (ranked/trusted/privatized) with respect to factors affecting the environment and their dependant measures. Collective Confidence Modeling Calculating the confidence value for a specific user with respect to an item/items, bearing in mind overall derived confidence from others users at the same time (collective notion). Modeling Confidence

Individual Confidence Modeling:

Calculating confidence for a specific user, with respect to

item/items which user has (ranked/trusted/privatized) with

respect to factors affecting the environment and their

dependant measures.

Collective Confidence Modeling

Calculating the confidence value for a specific user with

respect to an item/items, bearing in mind overall derived

confidence from others users at the same time (collective

notion).

Crowds of individuals are wise (Collective Wisdom) “ Aggregation of information in groups, resulting in decisions that are often better than could have been made by any single member of the group. “ Types of Crowds: Agent Crowds (Reputation Mechanisms) Fuzzy trust evaluation and credibility development in multi-agent systems , S Schmidt, R Steele, TS Dillon, E Chang - Applied Soft Computing Journal, 2007 – Elsevier User Crowds (Collaborative Filtering) Collecting community wisdom: integrating social search & social navigation , J Freyne, R Farzan, P Brusilovsky, B Smyth, … Wisdom of Crowds

Crowds of individuals are wise (Collective Wisdom)

“ Aggregation of information in groups, resulting in decisions

that are often better than could have been made by any single

member of the group. “

Types of Crowds:

Agent Crowds (Reputation Mechanisms)

Fuzzy trust evaluation and credibility development in multi-agent systems , S Schmidt, R Steele, TS Dillon, E Chang - Applied Soft Computing Journal, 2007 – Elsevier

User Crowds (Collaborative Filtering)

Collecting community wisdom: integrating social search &

social navigation , J Freyne, R Farzan, P Brusilovsky, B Smyth,

…

Fuzzification (pre-processing) phase: Applying fuzzification methodology to crisp values Defining membership functions Applying fuzzy rules Defuzzification (post-processing) phase: Feeding input values to membership functions Applying defuzzification methodology to fuzzy sets Calculating the Confidence

Fuzzification (pre-processing) phase:

Applying fuzzification methodology to crisp values

Defining membership functions

Applying fuzzy rules

Defuzzification (post-processing) phase:

Feeding input values to membership functions

Applying defuzzification methodology to fuzzy sets

Q Privacy Fuzzifier Trust Fuzzifier Rating Fuzzifier

Experiment setting Simulated 100 weighted user profiles Users have (virtually) visited two Smartmuseums, One located in Valletta, Malta. One located in Florence, Italy. selected records pertaining to users' profiles are taken and processed according to our confidence model Profile Records contain values pertaining to 7 cultural artifacts Two Maltese Three Florentine

Simulated 100 weighted user profiles

Users have (virtually) visited two Smartmuseums,

One located in Valletta, Malta.

One located in Florence, Italy.

selected records pertaining to users' profiles are taken and processed according to our confidence model

Profile Records contain values pertaining to 7 cultural artifacts

Two Maltese

Three Florentine

Modeling Sparsity Crisp Privacy Values

Filled Radar presentation of individual confidence Values

Filled Radar presentation of Collective confidence Values

Comparison of Standard Error Deviation

Comparison between input values with resulting confidence degrees Results are not uniform Comparison between pure confidence values and collective confidence CCF values are more uniformly distributed over diagram in comparison to pure confidence degrees and values. We realize that considering collective opinions while evaluating the confidence of an individual user over a certain item, could give better results With respect to uniformity (analytical perspective). Analysis

Comparison between input values with resulting confidence degrees

Results are not uniform

Comparison between pure confidence values and collective confidence

CCF values are more uniformly distributed over diagram in

comparison to pure confidence degrees and values.

We realize that considering collective opinions while evaluating the confidence of an individual user over a certain item, could give better results

With respect to uniformity (analytical perspective).

We introduced a fuzzy confidence model a fuzzy approach to modeling and analyzing confidence based on weights assigned to profiled information of users stored in semantic profiles. Based on our approach weights can be processed through a fuzzy reasoner and create a weighted outcome based on factors affecting the context of calculation. Conclusion

We introduced a fuzzy confidence model

a fuzzy approach to modeling and analyzing confidence based

on

weights assigned to profiled information of users stored in semantic profiles.

Based on our approach weights can be processed

through a fuzzy reasoner and create a weighted outcome based on factors affecting the context of calculation.

Experiment: We tested our approach with data from a real-world scenario where exhibitors of visual art experience personalized services of distributed knowledge-platforms. Novelty : We introduced a classic and a collective notion of confidence where values could be used to improve quality of adaptive personalized services allow us to detect similar individual or group behavioral patterns Conclusion

Experiment: We tested our approach

with data from a real-world scenario

where exhibitors of visual art experience personalized services of distributed knowledge-platforms.

Novelty :

We introduced a classic and a collective notion of confidence

where values could be used to improve quality of adaptive personalized services

allow us to detect similar individual or group behavioral patterns

Questions ? Thank you! Nima Dokoohaki Software and Computer Systems ( SCS ), Department of Electronics ,Computer and Software Systems ( ECS ), School of Information and Communications Technology ( ICT ), Royal Institute of Technology ( KTH ), Stockholm, Sweden Office: +46 (0) 8 790 4149 Cell : +46 (0) 76 269 76 30 Fax: +46 (0) 8 751 1793 Email : nimad@kth.se http://web.it.kth.se/~nimad/

Thank you!

Nima Dokoohaki

Software and Computer Systems ( SCS ), Department of Electronics ,Computer and

Software Systems ( ECS ), School of Information and Communications Technology ( ICT ), Royal Institute of Technology ( KTH ), Stockholm, Sweden Office: +46 (0) 8 790 4149 Cell : +46 (0) 76 269 76 30 Fax: +46 (0) 8 751 1793 Email : nimad@kth.se

http://web.it.kth.se/~nimad/