Feature selection
A novel method and/or system of feature selection is described.
1. A method for a filter-type feature selection for on-line content using a map reduce-type computing network, the method comprising:
executing computer instructions on at least one processor of at least one computing device that includes at least one memory, the computer instructions having been stored in the at least one memory, the computer instructions comprising:
generating, with the at least one processor, from a set of signal samples a set of feature signal vectors for a set of features for the on-line content in which a respective signal vector in the set of feature signal vectors corresponds to a respective feature in the set of features;
determining at least the following for the respective feature signal vectors in the set of feature signal vectors:
a normalized version of Hilbert-Schmidt Independence between a respective feature signal vector of the set of feature signal vectors and an output signal array,
a normalized version of Hilbert-Schmidt Independence between the respective feature signal vector of the set of feature signal vectors and the remaining feature signal vectors of the set of feature signal vectors, and
a gradient per feature of the set of features for the on-line content using the previously determined normalized versions of Hilbert-Schmidt Independence for the set of feature signal vectors;
selecting a feature having the feature signal vector of the set of feature signal vectors that produces the largest gradient;
storing the selected feature in the at least one memory; and
comparing, via accessing the at least one memory with the at least one processor and on a feature-by-feature basis, feature similarity between particular content on a media network coupled to the at least one computing device and the on-line content to determine a set of non-redundant features for the particular content based, at least in part, on the selected feature.
2. The method of claim 1 , and further comprising:
removing the signal samples from the set of signal samples and the feature signal vector from the set of feature signal vectors associated with the selected feature to produce an updated set of feature signal vectors;
determining the following for the respective feature signal vectors in the set of updated feature signal vectors:
a normalized version of Hilbert-Schmidt Independence between a respective feature signal vector of the set of updated feature signal vectors and an output signal array,
a normalized version of Hilbert-Schmidt Independence between the respective feature signal vector of the set of updated feature signal vectors and the remaining feature signal vectors of the set of updated feature signal vectors, and
a gradient per feature of the set of features using the previously determined normalized versions of Hilbert-Schmidt Independence for the set of updated feature signal vectors;
selecting a feature having the feature signal vector of the updated set of feature signal vectors that produces the largest gradient; and
repeating the removing, determining and selecting until a particular number of features are selected.
3. The method of claim 1 , wherein the gradient is computed in accordance with a least-angle regression (LARS) process.
4. The method of claim 1 , wherein the determining comprises using the Nystrom approximation to determine a normalized version of Hilbert-Schmidt Independence between a respective feature signal vector of the set of updated feature signal vectors and an output signal array, and to determine a normalized version of Hilbert-Schmidt Independence between the respective feature signal vector of the set of updated feature signal vectors and the remaining feature signal vectors of the set of updated feature signal vectors.
5. The method of claim 4 , wherein the determining the normalized version of Hilbert-Schmidt Independence and the determining the gradient is computed in a distributed computing manner.
6. The method of claim 2 , wherein the determining comprises using the Nystrom approximation to determine a normalized version of Hilbert-Schmidt Independence between a respective feature signal vector of the set of updated feature signal vectors and an output signal array, and to determine a normalized version of Hilbert-Schmidt Independence between the respective feature signal vector of the set of updated feature signal vectors and the remaining feature signal vectors of the set of updated feature signal vectors.
7. The method of claim 6 , wherein the determining the normalized version of Hilbert-Schmidt Independence and the determining the gradient is computed in a distributed computing manner.
8. The method of claim 7 , wherein the method is executed on a distributed computing architecture system.
9. The method of 1 , wherein the feature selected comprises a non-redundant feature.
10. The method of claim 9 , wherein the size of the signal sample set exceeds 20,000 samples.
11. The method of claim 10 , wherein the dimensionality of the signal sample set exceeds 1000.
12. The method of claim 11 , wherein the number of features selected exceeds 50.
13. The method of claim 11 , wherein the signal sample set comprises a signal sample set of features having non-zero-values.
14. A system for a filter-type feature selection for on-line content using a map reduce-type computing network, the system comprising:
a computing platform comprising at least one computing device
including at least one processor and at least one memory;
the at least one computing device to execute computer instructions on the at least one processor;
the computer instructions to be executed to have been stored on the at least one memory for execution on the at least one processor, the computer instructions comprise instructions to:
generate from a set of signal samples a set of feature signal vectors for a set of features for the on-line content in which a respective signal vector in the set of feature signal vectors corresponds to a respective feature in the set of features;
determine at least the following for the respective feature signal vectors in the set of feature signal vectors:
a normalized version of Hilbert-Schmidt Independence between a respective feature signal vector of the set of feature signal vectors and an output signal array,
a normalized version of Hilbert-Schmidt Independence between the respective feature signal vector of the set of feature signal vectors and the remaining feature signal vectors of the set of feature signal vectors, and
a gradient per feature of the set of features for the on-line content to be determined with the previously determined normalized versions of Hilbert-Schmidt Independence for the set of feature signal vectors; and
select a feature having the feature signal vector of the set of feature signal vectors to produce the largest gradient;
store the selected feature in the at least one memory; and
compare, via an access to the at least one memory with the at least one processor and on a feature-by-feature basis, feature similarity between particular content on a media network coupled to the at least one computing device and the on-line content to determine a set of non-redundant features for the particular content based, at least in part, on the selected feature.
15. The system of claim 14 , wherein the computer instructions further to:
remove the signal samples from the set of signal samples and the feature signal vector from the set of feature signal vectors associated with the selected feature to produce an updated set of feature signal vectors;
determine the following for the respective feature signal vectors in the set of updated feature signal vectors:
a normalized version of Hilbert-Schmidt Independence between a respective feature signal vector of the set of updated feature signal vectors and an output signal array,
a normalized version of Hilbert-Schmidt Independence between the respective feature signal vector of the set of updated feature signal vectors and the remaining feature signal vectors of the set of updated feature signal vectors, and
a gradient per feature of the set of features to be determined with the previously determined normalized versions of Hilbert-Schmidt Independence for the set of updated feature signal vectors; and
select a feature having the feature signal vector of the updated set of feature signal vectors to produce the largest gradient.
16. The system of claim 15 , wherein the computer instructions further to again remove, determine and select one or more additional times until a particular number of features is reached.
17. The system of claim 14 , wherein the system comprises a distributed computing network, wherein the computing platform comprises more than one computing device in the distributed computing network, and wherein the more than one computing devices in the distributed computing network are to determine for the respective feature signal vectors in the set of feature signal vectors in a distributed computing fashion.
18. A system for a filter-type feature selection for on-line content using a map reduce-type computing network, the system comprising:
means for generating, with the at least one processor, from a set of signal samples a set of feature signal vectors for a set of features for the on-line content in which a respective signal vector in the set of feature signal vectors corresponds to a respective feature in the set of features;
means for determining at least the following for the respective feature signal vectors in the set of feature signal vectors:
a normalized version of Hilbert-Schmidt Independence between a respective feature signal vector of the set of feature signal vectors and an output signal array,
a normalized version of Hilbert-Schmidt Independence between the respective feature signal vector of the set of feature signal vectors and the remaining feature signal vectors of the set of feature signal vectors, and
a gradient per feature of the set of features for the on-line content using the previously determined normalized versions of Hilbert-Schmidt Independence for the set of feature signal vectors;
means for selecting a feature having the feature signal vector of the set of feature signal vectors that produces the largest gradient;
means for storing the selected feature in at least one memory; and
means for comparing, via accessing the at least one memory with the at least one processor and on a feature-by-feature basis, feature similarity between particular content on a media network coupled to the at least one computing device and the on-line content to determine a set of non-redundant features for the particular content based, at least in part, on the selected feature.
19. The system of claim 18 , and further comprising:
means for removing the signal samples and feature signal vector associated with the selected feature to produce an updated set of feature signal vectors;
means for determining the following for the respective feature signal vectors in the set of updated feature signal vectors:
a normalized version of Hilbert-Schmidt Independence between a respective feature signal vector of the set of updated feature signal vectors and an output signal array,
a normalized version of Hilbert-Schmidt Independence between the respective feature signal vector of the set of updated feature signal vectors and the remaining feature signal vectors of the set of updated feature signal vectors, and
a gradient per feature of the set of features using the previously determined normalized versions of Hilbert-Schmidt Independence for the set of updated feature signal vectors; and
means for selecting a feature having the feature signal vector of the updated set of feature signal vectors that produces the largest gradient.
20. The system of claim 19 , and further comprising means for repeating removing, determining and selecting until a particular number of features are selected.