Generating a feature set
View Patent ↗A technique to generate a feature set. A plurality of samples from a data set can be clustered. Features can be selected based on the clusters. The features can be added to the feature set. Additional samples can be clustered and features selected and added to the feature set until a convergence threshold is reached.
1. A method executed by a system comprising a processor and a memory, comprising:
generating a feature set having a smaller number of features than features of a feature space corresponding to feature vectors representing data objects of a data set, the generating comprising using a plurality of iterations of different samplings of samples from the data set, the plurality of iterations comprising:
in a first iteration:
clustering a first subset of samples from the data set into a first plurality of clusters using a clustering algorithm;
selecting a first plurality of features based on the first plurality of clusters;
adding the first plurality of features to the feature set;
in each subsequent iteration following the first iteration:
clustering an additional plurality of samples from the data set, the additional plurality of samples being in addition to samples sampled from the data set in a previous iteration of the plurality of iterations;
selecting additional features based on resulting clusters from clustering the additional plurality of samples;
adding the additional features to the feature set;
continuing with the plurality of iterations until a convergence threshold is reached; and
clustering the data set using the clustering algorithm and the feature set, the clustering of the data set using the clustering algorithm and the feature set using a number of feature vectors that are simultaneously fit into the memory, the number of feature vectors representing respective data objects of the data set.
2. The method of claim 1 , further comprising:
generating the feature space by performing a term frequency-inverse document frequency (TF-IDF) analysis of the data set to reduce a dimensionality relative to an original feature space associated with the data set.
3. The method of claim 1 , wherein the convergence threshold is reached responsive to growth of the feature set in the plurality of iterations exhibiting a falling profile.
4. The method of claim 1 , wherein the convergence threshold is reached responsive to the number of features added to the feature set during an iteration being less than a threshold.
5. The method of claim 1 , wherein the convergence threshold is reached responsive to a number of plurality of iterations being greater than a threshold.
6. The method of claim 1 , wherein the first subset of samples and the additional plurality of samples are independent, random samples from the data set.
7. The method of claim 1 , wherein features are selected by:
ranking features in each cluster by information gain with respect to cluster inclusion; and
identifying top N ranked features from each cluster.
8. The method of claim 1 , wherein features are added to the feature set only if they are not already present in the feature set.
9. The method of claim 1 , wherein each feature vector including features of the feature set is smaller in size than a feature vector including features of the feature space.
10. The method of claim 1 , wherein the feature set approximates an ideal feature set for the feature space.
11. A system, comprising:
a processor;
a memory; and
a non-transitory storage medium storing instructions executable on the processor to:
generate a feature set having a smaller number of features than features of a feature space corresponding to feature vectors representing data objects of a data set, the generating comprising using a plurality of iterations of different samplings of samples from the data set, and aggregating features selected based on multiple clusterings of multiple pluralities of samples from the data set in the plurality of iterations until a convergence threshold is reached, the plurality of iterations comprising:
in a first iteration:
cluster a first subset of samples from the data set into a first plurality of clusters using a clustering algorithm;
select a first plurality of features based on the first plurality of clusters;
add the first plurality of features to the feature set;
in each subsequent iteration following the first iteration:
cluster an additional plurality of samples from the data set, the additional plurality of samples being in addition to samples sampled from the data set in a previous iteration of the plurality of iterations;
select additional features based on resulting clusters from clustering the additional plurality of samples;
add the additional features to the feature set; and
cluster the data set using the clustering algorithm and the feature set, using a number of feature vectors that are simultaneously fit into the memory, the number of feature vectors representing respective data objects of the data set and comprising features of the feature set that are less in number than the features of the feature space.
12. The system of claim 11 , wherein the first subset of samples and the additional plurality of samples are independent, random samples of the data set for use in generating the feature set.
13. The system of claim 11 , wherein each feature vector including features of the feature set is smaller in size than a feature vector including features of the feature space.
14. The system of claim 11 , wherein the feature set approximates an ideal feature set for the feature space.
15. A non-transitory computer readable storage medium storing instructions that when executed cause a system to:
generate a feature set having a smaller number of features than features of a feature space corresponding to feature vectors representing data objects of a data set, the generating comprising using a plurality of iterations of different samplings of samples from the data set, the plurality of iterations comprising:
in a first iteration:
cluster a first subset of samples from the data set into a first plurality of clusters using a clustering algorithm;
select a first plurality of features based on the first plurality of clusters:
add the first plurality of features to the feature set;
in each subsequent iteration following the first iteration:
cluster an additional plurality of samples from the data set, the additional plurality of samples being in addition to samples sampled from the data set in a previous iteration of the plurality of iterations;
select additional features based on resulting clusters from clustering the additional plurality of samples;
add the additional features to the feature set;
continue with the plurality of iterations until a convergence threshold is reached;
cluster the data set using the clustering algorithm and the feature set, using a number of feature vectors that are simultaneously fit into the memory, the number of feature vectors representing respective data objects of the data set and comprising features of the feature set that are less in number than the features of the feature space.
16. The non-transitory computer readable storage medium of claim 15 , wherein the convergence threshold is reached responsive to a number of features added to the feature set for clustering in a given iteration being below a threshold.
17. The non-transitory computer readable storage medium of claim 15 , wherein the different samplings comprise samplings using a random sampling algorithm.
18. The non-transitory computer readable storage medium of claim 15 , wherein each feature vector including features of the feature set is smaller in size than a feature vector including features of the feature space.
19. The non-transitory computer readable storage medium of claim 15 , wherein the feature set approximates an ideal feature set for the feature space.