exploiting context analysis for combining multiple entity resolution systems

21
Exploiting Context Analysis for Combining Multiple Entity Resolution Systems -Ramu Bandaru Zhaoqi Chen Dmitri V.kalashnikov Sharad Mehrotra

Upload: joshua

Post on 12-Jan-2016

55 views

Category:

Documents


0 download

DESCRIPTION

Exploiting Context Analysis for Combining Multiple Entity Resolution Systems. Zhaoqi Chen Dmitri V.kalashnikov Sharad Mehrotra. -Ramu Bandaru. Table of Contents. Introduction Related Work Problem Definition - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Exploiting Context Analysis for Combining Multiple Entity Resolution Systems

Exploiting Context Analysis for Combining Multiple Entity

Resolution Systems

-Ramu Bandaru

Zhaoqi Chen Dmitri V.kalashnikov Sharad Mehrotra

Page 2: Exploiting Context Analysis for Combining Multiple Entity Resolution Systems

Table of Contents

•Introduction•Related Work•Problem Definition•Context based Framework•Experimental Evaluation•Conclusion

Page 3: Exploiting Context Analysis for Combining Multiple Entity Resolution Systems

Introduction• This paper proposes a new Entity Resolution

Ensemble framework.• The real world raw datasets are often not perfect

and contain various types of issues, including errors and uncertainty, which complicate analysis.

• The task of ER Ensemble is to combine the results of multiple base-level ER systems into a single solution with the goal of increasing the quality of ER.

• The goal of ER is to identify and group references that

co-refer. The output of an ER system is a clustering of references, where each cluster is supposed to represent one distinct entity.

Page 4: Exploiting Context Analysis for Combining Multiple Entity Resolution Systems

• There are diverse ER approaches have been proposed to address the entity resolution challenge. This motivates the problem of creating a single combined ER solution – an ER ensemble.

• Different ER solutions perform better in different contexts. No solution is the best in terms of Quality.

• ER ensemble actively uses the notion of the correct clustering A+ for D.

• Each system Si is applied to the dataset being processed to produce its clustering A(i).

• Clustering A(i) corresponds to what system Si believes A+ should be.

Page 5: Exploiting Context Analysis for Combining Multiple Entity Resolution Systems

• The goal of ER ensemble is to build a clustering A* with the highest quality which is as close to A+ as possible.

• Difference between the two ensemble problems:• The objective of the cluster ensemble will be to

find a clustering that agrees the most with the n clusterings, including the n−1 poor-quality clusters.

• The goal of the ER ensemble is to simply output A(n) as its A*,and completely ignore the rest of the clusterings.▫ Let K be the true number of clusters

if (K is large) then high thresholdif (K is small) then low threshold

Page 6: Exploiting Context Analysis for Combining Multiple Entity Resolution Systems

• Two novel ER ensemble approaches. •A predictive way of selecting context features

that relies on estimating the unknown parameters

of the data being processed. •An extensive empirical evaluation of the

proposed solution.

Page 7: Exploiting Context Analysis for Combining Multiple Entity Resolution Systems

Related Work

• Entity Resolution : Many of the approaches have focused on exploiting various similarity metrics, relational similarities, probabilistic models.

• Clustering Ensembles : The final clustering is usually the one that agrees the most with the given clusterings.

• Combining Classifiers: • Majority Voting-popular rule • Bagging & Boosting-combine classifiers of same

type. • Stacking is to perform cross-validation on the base-level dataset and create a meta-level dataset from the predictions of base-level classifiers.

Page 8: Exploiting Context Analysis for Combining Multiple Entity Resolution Systems

Problem DefinitionEntity Resolution

• Two references ri and rj are said to co-refer, denoted as ri ~ rj , if they refer to the same object, that is, if ori =

orj.• The co-reference set Cr for r is the set of all references

that co-refer with r ie., Cr = {q belongs to R:q~r}.• The output of an ER system S applied to dataset D is a

set of clusters A = {A1,A2, . . . ,A|A|}.• To assess the quality of the resulting clusterings, we

evaluate how different the clustering A is from the ground truth A+.

Page 9: Exploiting Context Analysis for Combining Multiple Entity Resolution Systems

ER Ensemble

• These systems can be applied to dataset D to cluster the set of references R. Each system Si will produce its set of clusters A(i) = {A1

(i) ,A(i)2 , . . . ,A(i)

|Ai|}. • The goal is to be able to combine the resulting

clusterings A(1),A(2), . . . ,A(n) into a single clustering A* such that the quality of A* is higher than that of A(i), for i = 1, 2, . . . , n.

• The task of ER ensemble is for each ej belongs to E to provide a mapping dj->a*

j . Here, a*j belongs {−1,

1} is the prediction of the combined algorithm for edge ej , where a*

j = 1 if the overall algorithm believes a+

j = 1, and a*j =−1 otherwise.

Page 10: Exploiting Context Analysis for Combining Multiple Entity Resolution Systems
Page 11: Exploiting Context Analysis for Combining Multiple Entity Resolution Systems

Problem Definition• Using Blocking to Improve Efficiency : Virtual

Connected Subgraph(VCS).• Blocking is important to improve the efficiency

and also the quality of the algorithms.

• Naïve Solution: Voting – It will count predictions.• Advanced Voting Scheme is Weighted

Voting(WV).• Problem with voting is that it is a static non-

adaptive scheme with respect to the context.

Page 12: Exploiting Context Analysis for Combining Multiple Entity Resolution Systems

Context Based Framework• Context Features : Effectiveness and Generality• Number of Clusters: This is derived from the

expected number of clusters per VCS as well as the number of clusters generated by each base-level system.

• Node Fanout: It exploits the nodes adjacent to the edges at the edge level.

• The overall context feature vector fj is computed as a linear combination of the feature vectors

• For m features,

• Adding more context features can increase the robustness and accuracy of the system

Page 13: Exploiting Context Analysis for Combining Multiple Entity Resolution Systems

Overview of the Approach

Page 14: Exploiting Context Analysis for Combining Multiple Entity Resolution Systems

Meta-level Classification• 2 algorithms for meta-level classification:• Context-Extended Classification: Uses Regression

to find K* and Nv*.The computes fji.

• Now it is possible to learn a mapping from features dj ×fj into edge predictions aj

* by training a classifier using the ground truth labels aj

+ .

Page 15: Exploiting Context Analysis for Combining Multiple Entity Resolution Systems

• Context-Weighted Classification: The idea of this approach is to learn for each base-level system Si a model Mi that would predict how accurate Si tends to be in a given context.

Page 16: Exploiting Context Analysis for Combining Multiple Entity Resolution Systems

Creating Final Clusters

• Correlation Clustering (CC) generates the partitioning of data points that optimizes a score function.

• CC was specifically designed for the situation where there are conflicts in edge labeling.

• Input: A graph where nodes correspond to references and edges between two nodes are labeled, in the simplest case, with +1 if there is evidence that u and v co-refer, and with −1 otherwise.

• CC finds a clustering that has the most agreement with the edge labeling, therefore limiting the conflicts as much as possible.

Page 17: Exploiting Context Analysis for Combining Multiple Entity Resolution Systems

Experimental Evaluation• Set Up: Dataset, Evaluation Metrics, Baseline

Methods and Classifiers.• Experiments on Web Domain: Base-level

Algorithms, Comparison of various combination algorithms.

• The main difference among these algorithms are (a) the similarity metrics chosen to compare

pairs of references and (b) the clustering algorithms that

generate the final clusters based on similarities.

Page 18: Exploiting Context Analysis for Combining Multiple Entity Resolution Systems

• Efficiency: The running time of the ER ensemble algorithm consists of several parts: running base-level ER systems and loading the

decisions by these systems on the edges, running the two regression classifiers to derive the context features, applying meta-level classification to predict the edge classes, and creating final clusters.

• Effects of Blocking: Blocking is important to improve

the efficiency of our algorithms.• It also has great impact on the quality of ER

ensemble since the VCS level context features rely on the blocking.

Page 19: Exploiting Context Analysis for Combining Multiple Entity Resolution Systems

Conclusion• This paper develops a novel ER ensemble

framework for combining multiple base-level entity resolution

systems. • The Context-Extended and Context-Weighted

Ensemble approaches for meta-level classification that enable the framework to adapt to the dataset being processed based on the local context is also proposed.

• This study demonstrates the superiority of the Context-Weighted Ensemble approach.

Page 20: Exploiting Context Analysis for Combining Multiple Entity Resolution Systems

Future Work of the Paper• To explore principally different ways to create

context features and a different classification scheme that trades quality for efficiency.

Page 21: Exploiting Context Analysis for Combining Multiple Entity Resolution Systems

Thank You!!!