幻灯片 1 - Artificial Intelligence Laboratory

sounderslipInternet and Web Development

Oct 22, 2013 (4 years and 2 months ago)

98 views

Conferences Review



AAAI and IJCAI


Sean

2

Outline


Introduction to AAAI



Selected papers from AAAI (3)



Introduction to IJCAI



Selected papers from IJCAI (3)



Summary

3

Introduction to AAAI


Association for the Advancement of Artificial Intelligence conference
on Artificial Intelligence (AAAI)


Annual conference in summer (from 1980)


Totally 24 sessions by now


Acceptance rate: 25%~30%


No AAAI 2009



Related tracks


AI and the Web Track (Special track)


Natural Language Processing


Knowledge
-
Based Information Systems


Machine Learning



Major groups are from engineering school (algorithms and IS)


Qiang Yang et al., HKUST, Hong Kong


Changshui Zhang et al., Tsinghua University, China


Zhejiang University, China


Zhi
-
Hua Zhou et al. Nanjing University, China

4

Selected Papers from AAAI


AAAI
-
10 outstanding paper awards


How Incomplete Is Your Semantic Web Reasoner? Systematic
Analysis of the Completeness of Query Answering Systems


AI and the Web Track (Special track, AAAI
-
10)


Giorgos Stoilos, Bernardo Cuenca Grau, Ian Horrocks (Oxford
University, UK)



Other selected papers


Modeling Dynamic Multi
-
Topic Discussions in Online Forums


AI and the Web Track (Special track, AAAI
-
10)


Hao Wu, Jiajun Bu, Chun Chen, Can Wang, Guang Qiu, Lijun
Zhang and Jianfeng Shen (Zhengjiang University, China)



Learning to Predict Opinion Share in Social Networks


AI and the Web Track (Special track, AAAI
-
10)


Masahiro Kimura, Kazumi Saito, Kouzou Ohara, Hiroshi Motoda
(Osaka University et al., Japan)

5

Selected Papers from AAAI


AAAI
-
10 outstanding paper awards


How Incomplete Is Your Semantic Web Reasoner? Systematic
Analysis of the Completeness of Query Answering Systems


AI and the Web Track (Special track, AAAI
-
10)


Giorgos Stoilos, Bernardo Cuenca Grau, Ian Horrocks (Oxford
University, UK)



Other selected papers


Modeling Dynamic Multi
-
Topic Discussions in Online Forums


AI and the Web Track (Special track, AAAI
-
10)


Hao Wu, Jiajun Bu, Chun Chen, Can Wang, Guang Qiu, Lijun
Zhang and Jianfeng Shen (Zhengjiang University, China)



Learning to Predict Opinion Share in Social Networks


AI and the Web Track (Special track, AAAI
-
10)


Masahiro Kimura, Kazumi Saito, Kouzou Ohara, Hiroshi Motoda
(Osaka University et al., Japan)

6

Introduction


Introduction


Web Ontology Language (OWL) plays a key role in the Semantic Web
Reasoner of a query answering system


For data: describe the meaning of the data


For user: provide answers to query



Completeness vs. efficiency


Completeness: use ontology to provide all possible answers to a query


Efficiency: ignore ontology, just use simple matching


In practical applications, incompleteness is chosen, which lies between
completeness and efficiency



Research question and challenges


How to evaluate the completeness of a semantic web reasoner?


Data is not generic and exhaustive (to provide all possible answers to a
query)


Answers may not be verifiable

7

Algorithms


Ontology benchmark: Lehigh University Benchmark (LUBM)


An ontology describing an academic domain



Including the ontology, the testing datasets and testing queries



Proposed framework


Step 1: generate an “n
-
exhaustive” testing datasets based on LUBM
ontology using the proposed algorithm (SyGENiA)


The generation of testing datasets in LUBM are hard
-
coded and is not
exhaustive


Exhaustive testing datasets is proved to be impossible to generate due to
exponential increase of computing time w.r.t. the scale of the ontology


“n
-
exhaustive” testing datasets can be used as an approximation to
exhaustive testing datasets, which is derived by adding some constraints to
the generation of exhaustive testing datasets



Step 2: test the proposed “n
-
exhaustive” testing datasets generated by
SyGENiA using some query answering systems and compare the result
to that of the benchmark (LUBM)

8

Results


The results show that


For all 4 systems, the testing datasets generated by SyGENiA
indicate more incompleteness that of LUBM


Provide a practical algorithms to generate testing datasets to test
the completeness of query answering systems



For AI lab research


Build and test ontology for online text in social media (BI)

9

Selected Papers from AAAI


AAAI
-
10 outstanding paper awards


How Incomplete Is Your Semantic Web Reasoner? Systematic
Analysis of the Completeness of Query Answering Systems


AI and the Web Track (Special track, AAAI
-
10)


Giorgos Stoilos, Bernardo Cuenca Grau, Ian Horrocks (Oxford
University, UK)



Other selected papers


Modeling Dynamic Multi
-
Topic Discussions in Online Forums


AI and the Web Track (Special track, AAAI
-
10)


Hao Wu, Jiajun Bu, Chun Chen, Can Wang, Guang Qiu, Lijun
Zhang and Jianfeng Shen (Zhengjiang University, China)



Learning to Predict Opinion Share in Social Networks


AI and the Web Track (Special track, AAAI
-
10)


Masahiro Kimura, Kazumi Saito, Kouzou Ohara, Hiroshi Motoda
(Osaka University et al., Japan)

10

Introduction


Introduction


Topics diffuse among online social network by self
-
preference
and peer
-
influence



Three aspects to consider


The diffusion of generic information (B
-
TFM)


The diffusion of certain topics (T
-
TFM)


Fading of interest on topic during diffusion (TT
-
TFM)



Research questions


How to model the topic diffusion considering both self
-
preference
and peer
-
influence?



How to analyze the diffusion of specific topics at specific time?

11

Algorithms


B
-
TFM


Use reply
-
to relationship to build adjacent matrix of social
network for random walk (peer
-
influence)


Use the number of replies of each user to measure the intensity
of participation (self
-
preference)


Combine peer
-
influence and self
-
preference into a single
measure called “ParticipationRank”, updated at each time point



T
-
TFM


Use LDA for topic analysis of each thread


Build separate social networks for each topic, and use the topic
strength to adjust the transition probabilities in adjacent matrices



TT
-
TFM


Add time lapse factor such that the transition probabilities in the
adjacent matrix of each topic social network fade with time

12

Results


Dataset: Drag Racing, Honda/Acura (Honda
-
tech forum)



Task: to predict if a user will participate in the discussion of a specific topic
at a certain time point by ParticipationRank











Results show that TT
-
TFM performs the best



For AI lab research


Study viral marketing in social media (BI)

13

Selected Papers from AAAI


AAAI
-
10 outstanding paper awards


How Incomplete Is Your Semantic Web Reasoner? Systematic
Analysis of the Completeness of Query Answering Systems


AI and the Web Track (Special track, AAAI
-
10)


Giorgos Stoilos, Bernardo Cuenca Grau, Ian Horrocks (Oxford
University, UK)



Other selected papers


Modeling Dynamic Multi
-
Topic Discussions in Online Forums


AI and the Web Track (Special track, AAAI
-
10)


Hao Wu, Jiajun Bu, Chun Chen, Can Wang, Guang Qiu, Lijun
Zhang and Jianfeng Shen (Zhengjiang University, China)



Learning to Predict Opinion Share in Social Networks


AI and the Web Track (Special track, AAAI
-
10)


Masahiro Kimura, Kazumi Saito, Kouzou Ohara, Hiroshi Motoda
(Osaka University et al., Japan)

14

Introduction


Introduction


Multiple opinions diffusion in social network



Voter model



Value
-
weighted voter model




Property of value
-
weighted voter model


Eventually one opinion will win and others will die out


Share of opinion


The percentage of population that hold a certain opinion



Research questions


How to predict the share of opinions at a future time point in
social networks?

15

Algorithms


The algorithm aims to estimate the weight value of each opinion by
maximizing the log
-
likelihood function of the vector of weight values



Algorithm


Step 1: initialize all weight value to 0


Step 2: calculate the first order derivative of the log
-
likelihood function


Step 3: if the first order derivative is sufficiently small (below a given
threshold), terminate. Otherwise, go to step 4


Step 4: calculate the Hessian matrix (second order derivative) and
update the vector of weight values by multiplying the inverted Hessian
matrix, return to step 2



Benchmark


Naïve linear method: simple linear regression



Datasets (social networks)


Japanese blog networks, list of people in Japanese Wikipedia

16

Results








Results show that performance of predicting opinion
shares with the proposed learning method is better



For AI lab research


Topic/information diffusion in social media (BI/GeoPolitical)

17

Introduction to IJCAI


International Joint Conferences on Artificial Intelligence (IJCAI)


Biennial conference in summer (from 1969)


Totally 20 sessions by now


Acceptance rate: 20%~25%



Related tracks


Web and Knowledge
-
based Information Systems


Natural Language Processing


Machine Learning



Major groups are from engineering school (algorithms)


Changshui Zhang et al., Tsinghua University, China


Jieping Ye et al., Arizona State University, Arizona


Qiang Yang et al., HKUST, Hong Kong


Zhengjiang University, China


Zhi
-
Hua Zhou et al. Nanjing University, China


University of Illinois at Chicago, Illinois

18

Selected Papers from IJCAI


IJCAI
-
09 distinguished paper awards


Learning Conditional Preference Networks with Queries


Uncertainty in AI


Frédéric Koriche, Bruno Zanuttini (Université de Caen Basse
-
Normandie, France)



Other selected papers


Efficient Estimation of Influence Functions for SIS Model on
Social Networks


Web and Knowledge
-
based Information Systems


Masahiro Kimura, Kazumi Saito, Hiroshi Motoda (Osaka University
et al., Japan)



Incorporating User Behaviors in New Word Detection


Web and Knowledge
-
based Information Systems


Yabin Zheng, Zhiyuan Liu, Maosong Sun, Liyun Ru, Yang Zhang
(Tsinghua University, China)

19

Selected Papers from IJCAI


IJCAI
-
09 distinguished paper awards


Learning Conditional Preference Networks with Queries


Uncertainty in AI


Frédéric Koriche, Bruno Zanuttini (Université de Caen Basse
-
Normandie, France)



Other selected papers


Efficient Estimation of Influence Functions for SIS Model on
Social Networks


Web and Knowledge
-
based Information Systems


Masahiro Kimura, Kazumi Saito, Hiroshi Motoda (Osaka University
et al., Japan)



Incorporating User Behaviors in New Word Detection


Web and Knowledge
-
based Information Systems


Yabin Zheng, Zhiyuan Liu, Maosong Sun, Liyun Ru, Yang Zhang
(Tsinghua University, China)

20

Introduction


Introduction


Conditional Preference Networks (CP
-
nets)


A graph where each node (attribute) is labelled with a table describing the
user’s preference over alternative values of this node given different values
of the parent nodes



Traditional way of building CP
-
nets


Select possible attributes


Asking a user for the preference of each attribute


Build the CP
-
net by the collected information



Challenges


A minimal set of attributes must be selected to build the CP
-
net


Too many irrelevant attributes will lead to low efficiency



Research question


How to design an efficient algorithm to build CP
-
net by actively feeding
queries (preference relationships) to the algorithm?

21

Conditional Preference Networks

22

Algorithms

Test if a preference relationship is consistent in N (current CP
-
net)


If false, take the counter example


|

If there are rules (of a node) that involve the counter example


|

|

Find the parent nodes of the node


|

|

Expand the conditions of the rules using parent nodes


|

Else


|

|

Add the node and the rules to N


Return N



Advantages of the proposed algorithm


Integrates the learning and preference testing together, which are
separated in traditional way


The computational complexity is proved to be linear in the size of CP
-
net and logarithmic in the number of attributes



For AI lab research


Recommendation systems in social media (BI)

23

Selected Papers from IJCAI


IJCAI
-
09 distinguished paper awards


Learning Conditional Preference Networks with Queries


Uncertainty in AI


Frédéric Koriche, Bruno Zanuttini (Université de Caen Basse
-
Normandie, France)



Other selected papers


Efficient Estimation of Influence Functions for SIS Model on
Social Networks


Web and Knowledge
-
based Information Systems


Masahiro Kimura, Kazumi Saito, Hiroshi Motoda (Osaka
Universityet al., Japan)



Incorporating User Behaviors in New Word Detection


Web and Knowledge
-
based Information Systems


Yabin Zheng, Zhiyuan Liu, Maosong Sun, Liyun Ru, Yang Zhang
(Tsinghua University, China)

24

Introduction


Introduction


SIS (Susceptible
-
Infected
-
Susceptible) model describes the repeated diffusion of
a topic in social network



Influence function


σ(v,t): expected number of nodes infected by v at time t when v was infected at t=0



Research question


How to estimate the influence function of each node by effective (in terms of
computational time) simulation?



Layered graph method


All vertices are presented


Only edges through which topic diffused are added at time t


The graph (edges) evolve with time



Proposed technique and algorithms


Bond percolation (BP)


Bond percolation with pruning method: retain only one node when many nodes
have exactly the same influence path at time t

25

Algorithms

26

Results


Advantages


The influence function of all nodes are estimated simultaneously


The number of edges in the graph are significantly reduced when
propagation probability is small



For AI lab research


SIS/SIR model simulation in social media (BI/GeoPolitical)

27

Selected Papers from IJCAI


IJCAI
-
09 distinguished paper awards


Learning Conditional Preference Networks with Queries


Uncertainty in AI


Frédéric Koriche, Bruno Zanuttini (Université de Caen Basse
-
Normandie, France)



Other selected papers


Efficient Estimation of Influence Functions for SIS Model on
Social Networks


Web and Knowledge
-
based Information Systems


Masahiro Kimura, Kazumi Saito, Hiroshi Motoda (Osaka University
et al., Japan)



Incorporating User Behaviors in New Word Detection


Web and Knowledge
-
based Information Systems


Yabin Zheng, Zhiyuan Liu, Maosong Sun, Liyun Ru, Yang Zhang
(Tsinghua University, China)

28

Introduction


Introduction


Why study new word detection


Try to identify out
-
of
-
vocabulary words


Useful for language with no natural word boundaries (e.g. Chinese)



Lexicons


Cell dictionary: domain specific lexicons


User dictionary: user specific lexicons



Word features


Coverage: how many users have used a word (popularity)


Discriminability: the ratio of popularity of a word among users from a specific
domain and users outside that specific domain



Research question


How to detect new words in domain
-
specific fields based on user
behavior?

29

Algorithms


Step 1:


Identify top n representative words from every domain
using the combination of coverage and
discriminability



Step 2:


Identify users who use the representative words very
frequently as potential experts



Step 3:


Identify new words by their popularity among potential
experts and other users

30

Results


Dataset: generated from Sogou (
搜狗
) Chinese input method



Benchmarks: Google Sets, Bayesian Sets



Evaluation metrics (relevant documents)


Bpref: binary preference measure


MRR: mean reciprocal rank


P@n: precision at n









For AI lab research


Features selection for text mining in social media (BI)

31

Summary


All papers focused on algorithms
development



Possible take
-
away for AI lab


Topic diffusion analysis in social media for
both empirical analysis and simulation



Feature selection using collaborative filtering