1. The Painter’s Feature Selection for Gene Expression Data Lyon, 23-26 August 2007 Daniele Apiletti, Elena Baralis, Giulia Bruno, Alessandro Fiori

2. 2 Introduction Feature selection identifies a minimum set of relevant features is applied before a learning algorithm reduces computation costs increases the speed up of learning process increases the model interpretability improves the classification accuracy performance

3. 3 Framework Painter Microarray labeled data Feature Selection Gene Rank Model Building labels Classification Model Microarray unlabeled data Filter

4. 4 Feature selection From Painter’s Algorithm in Computer Graphics to paint shadows Overlap score to measure: common expression intervals in different classes gaps between expression intervals among classes Bonus to genes with largest gaps few overlapping classes gene expression value class1 class2 class3 class gene expression value 1 2 3 class 2 1 3 gap w1w1 w2w2 w4w4 w3w3 w5w5 w tot

5. 5 Example overlapscore =(1*6)+(1*4)/11=10/11 0 = no overlapping overlapscore= (1*1)+(2*4)+(1*1)/6=10/6 1 overlapping

6. 6 Filter Reason: noisy data outliers presence gene expression value

7. 7 Weighted interval gene expression value 35566304543

8. 8 Experimental design 10-cross validation SVM kernel: Crammer and Singer (CS) degree: 1 cost: 100 Method compared: analysis of variance (ANOVA) signal-to-noise ratio in OVO (OVO) signal-to-noise ratio in OVR (OVR) ratio of variables between categories to within categories sum of squares (BW) DatasetsPatientsGenesClasses Tumors96057279 Brain19059215 Brain260103644

9. 9 Experimental results (1) 200 genes selected PainterOVOBWANOVAOVR Brain186.7 ±8.888.9 ±7.486.7 ±9.188.89 ±6.387.8 ±7.0 Brain270.7 ±20.064.7 ±12.464.0 ±14.562.2 ±22.461,8 ±15.7 Tumors971.0 ±24.465.5 ±15.164.2 ±16.669.9 ±18.466.8 ±20.2 Average76.1 ±17.673.0±11.671.6 ±13.473.7 ±15.772.1 ±14.3

10. 10 Experimental results (2) trend on Brain2 dataset

11. 11 Conclusion New method: multi-class approach based on new criterion of gene relevance self adaptation to the datasets distribution density based filter smoothing outliers effect Results robustness of the algorithm can be applied to any dataset with continuous valued features Future work: investigation of features groups with the same distiminating power comparison with more feature selection techniques on other datasets

12. 12 Thanks for the attention!