Behaviour of FeRaNGA - Feature Ranking process using Inductive Modelling 0 100 1110 0 1100 10 1 0 1110 10 1 0 1110 0 10 Aleš Pilný 0 110 1111 0 1101110 0 110 1111 0 1110 110 pi l nya1@ f el . cvut . cz 0 110 0 0 0 1 0 0 100 0 0 0 0 1110 0 11 0 11010 11 0 1110 10 1 0 1110 0 0 0 Pavel Kordík, Miroslav Šnorek 0 110 10 0 1 0 1101110 kor di kp@ f el . cvut . cz, s nor ek@ f el . cvut . cz 0 110 0 0 0 1 0 0 100 0 0 0 0 110 10 11 0 1100 0 0 1 ht t p: //ci g. f el k. cvut . cz 0 1110 10 0 0 1100 10 1 0 110 0 10 0 0 1110 0 10 0 11110 0 1 0 0 100 0 0 0 0 1110 0 0 0 0 1101111 0 110 0 0 11 0 11010 0 1 Computational Intelligence Group 0 1110 10 0 0 1100 0 0 1 Department of Computer Science and Engineering 0 110 0 0 11 0 1110 10 1 Faculty of Electrical Engineering 0 0 10 110 0 0 0 100 0 0 0 0 100 0 110 0 10 00 10 1 Czech Technical University in Prague 0 100 110 0 0 0 100 0 0 0 ICANN 2008 0 100 0 0 11 0 10 10 110 0 1010 10 1 0 10 10 10 0
Overview of Feature Ranking and Selection How important is each feature? Ranks 1. P-length 2. P-width 3. S-length 4. S-width Feature Ranking Reduction Knowledge Of dimensionality Feature Selection ICANN 2008 Aleš Pilný, pilnya1@fel.cvut.cz, http://cig.felk.cvut.cz
The FAKE-GAME Tool overview ● Extension of MIA GMDH ICANN 2008 Aleš Pilný, pilnya1@fel.cvut.cz, http://cig.felk.cvut.cz
Feature Ranking(FR) in FAKE-GAME ● FAKE-GAME tool creates the GAME network using Niching Genetic Algorithm (NGA) ● Importance of each feature can be obtained as a side effect of NGA by computing utilization in net building process ● This approach also causes selection of important features by ignoring redundant and irrelevant features. ICANN 2008 Aleš Pilný, pilnya1@fel.cvut.cz, http://cig.felk.cvut.cz
Feature Ranking utilizing information from Niching Genetic Algorithm - FeRaNGA ● Novel approach for Feature Ranking ● Ranking is easily extracted from proportional significance of features ● How? – NGA = GA + domains (location of multiple solutions) – We used Deterministic Crowding method to promote the formation and maintenance of stable subpopulations in GA. – Significance is estimated by monitoring which genes exist in the population (which features are used by genes in NGA) ICANN 2008 Aleš Pilný, pilnya1@fel.cvut.cz, http://cig.felk.cvut.cz
FeRaNGA ● NGA random initialization → Problem with a results instability of FeRaNGA How to solve it? ICANN 2008 Aleš Pilný, pilnya1@fel.cvut.cz, http://cig.felk.cvut.cz
FeRaNGA-n ● NGA random initialization → Problem with a results instability of FeRaNGA ● All ranks are computed from ensemble of -n GAME models as a MEDIANS from estimated significance ICANN 2008 Aleš Pilný, pilnya1@fel.cvut.cz, http://cig.felk.cvut.cz
FeRaNGA-n ● NGA random initialization → Problem with a results instability of FeRaNGA ● All ranks are computed from ensemble of -n GAME models as a MEDIANS from estimated significance FeRaNGA-3 Correct ranks: FAKE-GAME models: Model 0: 1 2 3 5 4 1 2 3 5 4 1 2 3 4 5 Model 1: 1 3 2 4 5 Model 2: 1 2 3 4 5 ICANN 2008 Aleš Pilný, pilnya1@fel.cvut.cz, http://cig.felk.cvut.cz
Experiments 1.Influence of NGA configuration on ranks 2.Dependency of accuracy on Nr. of models for FeRaNGA-n method 3.Changes of ranks between layers Three kinds of experiments on two artificial data sets. ICANN 2008 Aleš Pilný, pilnya1@fel.cvut.cz, http://cig.felk.cvut.cz
The Data sets used in experiments ● Gaussian multivariate data set – two clusters of points generated from two different 10th- dimensional normal Gaussian distributions – 1-10 are equally relevant, 11-20 are irrelevant, 21-50 are highly redundant with the first ten features ● Uniform Hypercube data set – two clusters of points generated from two different 10th- dimensional hypercube [0 ; 1]¹º, with uniform distribution – 1-10 with decreasing relevance, 11-20 are irrelevant, 21-50 redundant ICANN 2008 Aleš Pilný, pilnya1@fel.cvut.cz, http://cig.felk.cvut.cz
1. Influence of NGA configuration on FeRaNGA-7 results (on Gaussian Data Set) Default configuration of NGA: 30 individuals and 15 epochs Ranks computed as a medians over all layers of medians. ICANN 2008 Aleš Pilný, pilnya1@fel.cvut.cz, http://cig.felk.cvut.cz
1. Influence of NGA configuration on FeRaNGA-7 results (on Gaussian Data Set) Default configuration of NGA: 30 individuals and 15 epochs 9 correct ranks in first two layers Redundant features Incorrect features! Ranks computed as a medians over all layers of medians ICANN 2008 Aleš Pilný, pilnya1@fel.cvut.cz, http://cig.felk.cvut.cz
1. Influence of NGA configuration on FeRaNGA-7 results (on Gaussian Data Set) Default configuration of NGA: 30 individuals and 15 epochs 9 correct ranks in first two layers Redundant features Incorrect features! Configuration of NGA: 75 individuals and 75 epochs 10 correct ranks ICANN 2008 Aleš Pilný, pilnya1@fel.cvut.cz, http://cig.felk.cvut.cz
1. Influence of NGA configuration on FeRaNGA-7 results (on Gaussian Data Set) Default configuration of NGA: 30 individuals and 15 epochs 9 correct ranks in first two layers Redundant features Incorrect features! Configuration of NGA: 75 individuals and 75 epochs 10 correct ranks Configuration of NGA: 150 individuals and 150 epochs All features have correct ranks. ICANN 2008 Aleš Pilný, pilnya1@fel.cvut.cz, http://cig.felk.cvut.cz
Dependency of accuracy on Nr. of models for FeRaNGA-n method (on Hypercube Data set) First ten ranks from first layers of FeRaNGA-7 on the Hypercube Data Set. ● Ranks computed from a higher Nr. of models depend on significance of features from previous models. ICANN 2008 Aleš Pilný, pilnya1@fel.cvut.cz, http://cig.felk.cvut.cz
Dependency of accuracy on Nr. of models for FeRaNGA-n method (on Hypercube Data set) First ten ranks from first layers of FeRaNGA-7 on the Hypercube Data Set. ICANN 2008 Aleš Pilný, pilnya1@fel.cvut.cz, http://cig.felk.cvut.cz
Dependency of accuracy on Nr. of models for FeRaNGA-n method (on Hypercube Data set) First ten ranks from first layers of FeRaNGA-7 on the Hypercube Data Set. ● For NGA configuration 75 are correct ranks from 5, 6 and 7 models. ICANN 2008 Aleš Pilný, pilnya1@fel.cvut.cz, http://cig.felk.cvut.cz
Dependency of accuracy on Nr. of models for FeRaNGA-n method (on Hypercube Data set) First ten ranks from first layers of FeRaNGA-7 on the Hypercube Data Set. ● For NGA configuration 75 are correct ranks from 5, 6 and 7 models. ● Growing Nr. of models and stronger NGA config. cause improving of accuracy. ICANN 2008 Aleš Pilný, pilnya1@fel.cvut.cz, http://cig.felk.cvut.cz
Dependency of accuracy on Nr. of models for FeRaNGA-n method (on Hypercube Data set) First ten ranks from first layers of FeRaNGA-7 on the Hypercube Data Set. ● For NGA configuration 75 are correct ranks from 5, 6 and 7 models. ● Growing Nr. of models and stronger NGA config. cause improving of accuracy. ● With NGA config. 150 are all ranks of features correct. ICANN 2008 Aleš Pilný, pilnya1@fel.cvut.cz, http://cig.felk.cvut.cz
Changes of ranks between layers (on Hypercube Data set) Changes of ranks between first two layers for 14 GAME models (cfg.150) ICANN 2008 Aleš Pilný, pilnya1@fel.cvut.cz, http://cig.felk.cvut.cz
Changes of ranks between layers (on Hypercube Data set) Changes of ranks between first two layers for 14 GAME models (cfg.150) ● In all cases the relevant features loses a part of their importance ICANN 2008 Aleš Pilný, pilnya1@fel.cvut.cz, http://cig.felk.cvut.cz
Changes of ranks between layers (on Hypercube Data set) Changes of ranks between first two layers for 14 GAME models (cfg.150) ● In all cases the relevant features loses a part of their importance ● The average loss on one relevant feature is -0,3. A gain on one redundant feature is 0,09 and a gain on one irrelevant feature is 0,07. (the numbers are relative to Nr. of features) ICANN 2008 Aleš Pilný, pilnya1@fel.cvut.cz, http://cig.felk.cvut.cz
Changes of ranks between layers (on Hypercube Data set) Changes of ranks between first two layers for 14 GAME models (cfg.150) ● In all cases the relevant features loses a part of their importance ● The average loss on one relevant feature is -0,3. A gain on one redundant feature is 0,09 and a gain on one irrelevant feature is 0,07. (the numbers are relative to Nr. of features) ● In first layer are ranked only a few most important features and in every next layer this important features loss its importance on behalf of redundant and irrelevant features. ICANN 2008 Aleš Pilný, pilnya1@fel.cvut.cz, http://cig.felk.cvut.cz
Conclusion ● Stronger NGA configuration causes better results but higher Nr. of epochs and individuals slow down a learning process. ● With growing Nr. of models is accuracy increasing. ● Power of FeRaNGA-n is in first layer where only a few important features are ranked and redundant and irrelevant features are unused. ICANN 2008 Aleš Pilný, pilnya1@fel.cvut.cz, http://cig.felk.cvut.cz
Questions Thank you for your attention. Any questions? pilnya1@fel.cvut.cz ICANN 2008 Aleš Pilný, pilnya1@fel.cvut.cz, http://cig.felk.cvut.cz
Recommend
More recommend