Limits...
Sparse Zero-Sum Games as Stable Functional Feature Selection.

Sokolovska N, Teytaud O, Rizkalla S, MicroObese consortiumClément K, Zucker JD - PLoS ONE (2015)

Bottom Line: Feature selection, therefore, can lead not only to a problem with a reduced dimensionality, but also reveal some knowledge on functional classes of variables.In this contribution, we propose a framework based on a sparse zero-sum game which performs a stable functional feature selection.In particular, the approach is based on feature subsets ranking by a thresholding stochastic bandit.

View Article: PubMed Central - PubMed

Affiliation: Institute of Cardiometabolism and Nutrition, ICAN, Assistance Publique Hôpitaux de Paris, Pitié-Salpêtrière Hospital, Paris, France; Sorbonne Universités, UPMC University Paris 6, UMR_S 1166, ICAN, NutriOmics Team, Paris, France; INSERM, UMR S U1166, NutriOmics Team, Paris, France.

ABSTRACT
In large-scale systems biology applications, features are structured in hidden functional categories whose predictive power is identical. Feature selection, therefore, can lead not only to a problem with a reduced dimensionality, but also reveal some knowledge on functional classes of variables. In this contribution, we propose a framework based on a sparse zero-sum game which performs a stable functional feature selection. In particular, the approach is based on feature subsets ranking by a thresholding stochastic bandit. We provide a theoretical analysis of the introduced algorithm. We illustrate by experiments on both synthetic and real complex data that the proposed method is competitive from the predictive and stability viewpoints.

No MeSH data available.


Accuracy of the stochastic bandit (EXP3) ranking as a function of ranked feature sets, the number of feature sets is 50.On the left: the number of active features equals 5; in the center: the number of kept parameters is 7, and on the right: the number of selected features is 10.
© Copyright Policy
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4556702&req=5

pone.0134683.g004: Accuracy of the stochastic bandit (EXP3) ranking as a function of ranked feature sets, the number of feature sets is 50.On the left: the number of active features equals 5; in the center: the number of kept parameters is 7, and on the right: the number of selected features is 10.

Mentions: On Fig 1 we show the 10-folds cross validation accuracy rate for the t-test feature selection. It is easy to see that the best performance is achieved when the number of kept parameters is the smallest, and the accuracy is around 0.8. Figs 2–5 demonstrate the performance for the stochastic bandits methods. Figs 2 and 4 show the accuracy for the ranking by the stochastic bandit without thresholding (EXP3); Figs 3 and 5 show the performance for the thresholding bandit. We plot the accuracy as a function of ranked feature sets. On each plot above on the left: the number of active features equals 5; above on the right: the number of kept parameters is 7, and below the number of selected features is 10. On each plot, we show the accuracies for the strategies of the obtained ranking, so that the first strategy has got the highest probability, and the last one has the lowest probability to be played. Figs 2 and 3 show accuracies for the case where the number of feature sets is 10; Figs 4 and 5 illustrate the performance for 50 feature sets. We observe that the stochastic bandits feature selection method ranks the features set quite well. We also see that the thresholding stochastic bandit outperforms the stochastic bandit, and finds more relevant sets of features.


Sparse Zero-Sum Games as Stable Functional Feature Selection.

Sokolovska N, Teytaud O, Rizkalla S, MicroObese consortiumClément K, Zucker JD - PLoS ONE (2015)

Accuracy of the stochastic bandit (EXP3) ranking as a function of ranked feature sets, the number of feature sets is 50.On the left: the number of active features equals 5; in the center: the number of kept parameters is 7, and on the right: the number of selected features is 10.
© Copyright Policy
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4556702&req=5

pone.0134683.g004: Accuracy of the stochastic bandit (EXP3) ranking as a function of ranked feature sets, the number of feature sets is 50.On the left: the number of active features equals 5; in the center: the number of kept parameters is 7, and on the right: the number of selected features is 10.
Mentions: On Fig 1 we show the 10-folds cross validation accuracy rate for the t-test feature selection. It is easy to see that the best performance is achieved when the number of kept parameters is the smallest, and the accuracy is around 0.8. Figs 2–5 demonstrate the performance for the stochastic bandits methods. Figs 2 and 4 show the accuracy for the ranking by the stochastic bandit without thresholding (EXP3); Figs 3 and 5 show the performance for the thresholding bandit. We plot the accuracy as a function of ranked feature sets. On each plot above on the left: the number of active features equals 5; above on the right: the number of kept parameters is 7, and below the number of selected features is 10. On each plot, we show the accuracies for the strategies of the obtained ranking, so that the first strategy has got the highest probability, and the last one has the lowest probability to be played. Figs 2 and 3 show accuracies for the case where the number of feature sets is 10; Figs 4 and 5 illustrate the performance for 50 feature sets. We observe that the stochastic bandits feature selection method ranks the features set quite well. We also see that the thresholding stochastic bandit outperforms the stochastic bandit, and finds more relevant sets of features.

Bottom Line: Feature selection, therefore, can lead not only to a problem with a reduced dimensionality, but also reveal some knowledge on functional classes of variables.In this contribution, we propose a framework based on a sparse zero-sum game which performs a stable functional feature selection.In particular, the approach is based on feature subsets ranking by a thresholding stochastic bandit.

View Article: PubMed Central - PubMed

Affiliation: Institute of Cardiometabolism and Nutrition, ICAN, Assistance Publique Hôpitaux de Paris, Pitié-Salpêtrière Hospital, Paris, France; Sorbonne Universités, UPMC University Paris 6, UMR_S 1166, ICAN, NutriOmics Team, Paris, France; INSERM, UMR S U1166, NutriOmics Team, Paris, France.

ABSTRACT
In large-scale systems biology applications, features are structured in hidden functional categories whose predictive power is identical. Feature selection, therefore, can lead not only to a problem with a reduced dimensionality, but also reveal some knowledge on functional classes of variables. In this contribution, we propose a framework based on a sparse zero-sum game which performs a stable functional feature selection. In particular, the approach is based on feature subsets ranking by a thresholding stochastic bandit. We provide a theoretical analysis of the introduced algorithm. We illustrate by experiments on both synthetic and real complex data that the proposed method is competitive from the predictive and stability viewpoints.

No MeSH data available.