Sign In
 
 
 
New ItemNew Item

Full name

Luca Nicotra 

Title

Joint Input-Output Substructure Mining for Graph Labeling 

Start Time

15:30 

Location

Gerace 

Abstract

Structured output learning typically deals with the cases that the output variables form a sequence or a tree, and employs dynamic programming to solve the related combinatorial problem. However the graph structure of many problems is beyond the reach of dynamic programming methods. In our thesis we introduce some structured output boosting approaches for graph labeling based on graph mining. Classification hypotheses are efficiently found in the joint space of input graph samples and output labels. To this aim we tested novel pruning conditions to avoid traversing infrequent or non-discriminative input-output graph patterns. To produce a prediction rule from hypotheses, linear programming boosting (LPBoost) is employed. The method is validated on a syntetic dataset and on protein-protein interfaces prediction and is compared to standard classification methods disregarding output relations.

Keywords

Machine Learning, Data Mining, Graph Mining, Bioinformatics, Structured Output Learning 

Supervisor(s)

Franco Turini 

Notes

 

Session

Attachments
slide-nicotra.pdf    
Created at 8/15/2009 10:23 AM  by  
Last modified at 9/10/2009 8:59 PM  by Cristian Dittamo