FRESH: a feature-extraction and feature-significance toolkit¶
Feature extraction and selection are vital components of many machine-learning pipelines. Here we outline an implementation of the FRESH (FeatuRe Extraction and Scalable Hypothesis testing) algorithm.
Feature extraction is the process of building derived, aggregate features from a time series dataset. The features created are designed to characterize the underlying time series in a way that is easier to interpret and often provides a more suitable input to machine-learning algorithms.
Following feature extraction, statistical significance tests between feature and target vectors can be applied. This allows selection of only those features with relevance (in the form of a p-value) as defined by the user.
Feature selection can improve the accuracy of a machine-learning algorithm by
- Simplifying the models used.
- Shortening the training time needed.
- Mitigating the curse of dimensionality.
- Reducing variance in the dataset to reduce overfitting.
Notebooks showing examples of the FRESH algorithm used in different applications can be found at KxSystems/ml/fresh/notebooks.
Load the FRESH library in isolation from the utilities section of the toolkit using
q)\l ml/ml.q q).ml.loadfile`:fresh/init.q
Data passed to the feature extraction procedure should contain an identifying (id) column, which groups the time-series into subsets from which features can be extracted. The id column can be inherent to the data or derived for a specific use-case (e.g. applying a sliding window onto the dataset).
Null values in the data should be replaced with derived values most appropriate to the column.
The feature extraction procedure supports columns of boolean, integer and floating-point types. Other datatypes should not be passed to the extraction procedure.
In particular, data should not contain text (strings or symbols), other than within the id column. If a text-based feature is thought to be important, one-hot, frequency or lexigraphical encoding can be used to convert the symbolic data to appropriate numerical values.
A range of formatting functions (e.g. null-filling and one-hot encoding) are supplied in the preprocessing section of the toolkit.
Feature extraction functions are defined in the script
fresh.q and found within the
|absenergy[x]||Sum of squares|
|abssumchange[x]||Absolute sum of the differences between successive datapoints|
|aggautocorr[x]||Aggregation (mean, median, variance and standard deviation) of an autocorrelation over all possible lags (1 - count[x])|
|agglintrend[x;chunklen]||Slope, intercept and rvalue for the series over aggregated max, min, variance or average for chunks of size
|augfuller[x]||Hypothesis test to check for a unit root in series|
|autocorr[x;lag]||Autocorrelation over specified
|binnedentropy[x;nbins]||Entropy of the series binned into
|c3[x;lag]||Measure of the non-linearity of the series lagged by
|changequant[x;ql;qh;isabs]||Aggregated value of successive changes within corridor specified by lower quantile
|cidce[x;isabs]||Measure of series complexity based on peaks and troughs in the dataset (boolean
|count[x]||Number of values within the series|
|countabovemean[x]||Number of values in the series with a value greater than the mean|
|countbelowmean[x]||Number of values in the series with a value less than the mean|
|eratiobychunk[x;numsegments]||Sum of squares of each region of the series split into
|firstmax[x]||Position of the first occcurance of the maximum value in the series relative to the series length|
|firstmin[x]||Position of the first occcurance of the minimum value in the series relative to the series length|
|fftaggreg[x]||Spectral centroid (mean), variance, skew, and kurtosis of the absolute Fourier-transform spectrum|
|hasdup[x]||Boolean value stating if the series contains any duplicate values|
|hasdupmax[x]||Boolean value stating if a duplicate of the maximum value exists in the series|
|hasdupmin[x]||Boolean value stating if a duplicate of the minimum value exists in the series|
|indexmassquantile[x;q]||Relative index such that
|kurtosis[x]||Adjusted G2 Fisher-Pearson kurtosis of the series|
|largestdev[x;ratio]||Boolean value stating if the standard deviation is
|lastmax[x]||Position of the last occcurance of the maximum value in the series relative to the series length|
|lastmin[x]||Position of the last occcurance of the minimum value in the series relative to the series length|
|lintrend[x]||Slope, intercept and r-value associated with the series|
|longstrikegtmean[x]||Length of the longest subsequence in the series greater than the series mean|
|longstrikeltmean[x]||Length of the longest subsequence in the series less than the series mean|
|max[x]||Maximum value of the series|
|mean[x]||Mean value of the series|
|meanabschange[x]||Mean over the absolute difference between subsequent series values|
|meanchange[x]||Mean over the difference between subsequent series values|
|mean2dercentral[x]||Mean value of the central approximation of the second derivative of the series|
|med[x]||Median value of the series|
|min[x]||Minimum value of the series|
|numcrossingm[x;crossval]||Number of crossings in the series over the value
|numcwtpeaks[x;width]||Number of peaks in the series following data smoothing via application of a Ricker wavelet of defined
|numpeaks[x;support]||Number of peaks in the series with a specified
|partautocorrelation[x;lag]||Partial autocorrelation of the series with a specified
|perrecurtoalldata[x]||Ratio of count of values occurring more than once to count of different values|
|perrecurtoallval[x]||Ratio of count of values occurring more than once to count of data|
|quantile[x;quantile]||The value of series greater than the
|rangecount[x;minval;maxval]||The number of values greater than or equal to
|ratiobeyondrsigma[x;r]||Ratio of values more than
|ratiovalnumtserieslength[x]||Ratio of number of unique values to total number of values|
|skewness[x]||Skew of the series indicating asymmetry within the series|
|spktwelch[x;coeff]||Cross power spectral density of the series at given
|stddev[x]||Standard deviation of series|
|sumrecurringdatapoint[x]||Sum of all points present in the series more than once|
|sumrecurringval[x]||Sum of all the values present within the series more than once|
|sumval[x]||Sum of values within the series|
|symmetriclooking[x]||Measure of symmetry in the series|
|treverseasymstat[x;lag]||Measure of asymmetry of the series based on
|valcount[x;val]||Number of occurrances of
|var[x]||Variance of the series|
|vargtstdev[x]||Boolean value stating if the variance of the dataset is larger than the standard deviation|
Feature extraction functions are not, typically, called individually.
Feature extraction involves applying a set of aggregations to subsets of the initial input data, with the goal of obtaining information that is more informative to the prediction of the target vector than the raw time series.
.ml.fresh.createfeatures function applies a set of aggregation functions to derive features. There are 57 such functions callable within the
.ml.fresh.feat namespace, although users may select a subset of these based on requirement.
Applies functions to subsets of initial data to create features
tis the input data in the form of a simple table.
aggsis the id column name (syms).
cnamesare the column names (syms) on which extracted features will be calculated (these columns should contain only numerical values).
ptabis a table containing the functions and parameters to be applied to the
cnamescolumns. This should be a modified version of
This returns a table keyed by id column and containing the features extracted from the subset of the data identified by the
q)m:30;n:100 q)10#tab:(date:raze m#'"d"$til n;time:(m*n)#"t"$til m;col1:50*1+(m*n)?20;col2:(m*n)?1f) date time col1 col2 --------------------------------------- 2000.01.01 00:00:00.000 1000 0.3927524 2000.01.01 00:00:00.001 350 0.5170911 2000.01.01 00:00:00.002 950 0.5159796 2000.01.01 00:00:00.003 550 0.4066642 2000.01.01 00:00:00.004 450 0.1780839 2000.01.01 00:00:00.005 400 0.3017723 2000.01.01 00:00:00.006 400 0.785033 2000.01.01 00:00:00.007 500 0.5347096 2000.01.01 00:00:00.008 600 0.7111716 2000.01.01 00:00:00.009 250 0.411597 q)show ptab:.ml.fresh.params / for documentation purposes this table has been truncated f | pnum pnames pvals valid ------------------------| ------------------------------------------------------ absenergy | 0 () () 1 abssumchange | 0 () () 1 count | 0 () () 1 countabovemean | 0 () () 1 countbelowmean | 0 () () 1 firstmax | 0 () () 1 firstmin | 0 () () 1 autocorr | 1 ,`lag ,0 1 2 3 4 5 6 7 8 9 1 binnedentropy | 1 ,`lag ,2 5 10 1 c3 | 1 ,`lag ,1 2 3 1 cidce | 1 ,`boolean ,01b 1 eratiobychunk | 1 ,`numsegments ,3 1 rangecount | 2 `minval`maxval -1 1 1 changequant | 3 `ql`qh`isabs (0.1 0.2;0.9 0.8;01b) 1 q)5#cfeats:.ml.fresh.createfeatures[tab;`date;2_ cols tab;ptab] date | col1_absenergy col1_abssumchange col1_count col1_countabovemean col1_countb.. ----------| ---------------------------------------------------------------------------.. 2000.01.01| 1.33e+07 10100 30 13 17 .. 2000.01.02| 1.023e+07 11450 30 14 16 .. 2000.01.03| 7805000 9200 30 13 17 .. 2000.01.04| 8817500 9950 30 17 13 .. 2000.01.05| 7597500 7300 30 12 18 .. q)count 1_cols cfeats / 595 features have been produced from 2 columns 568 / update ptab to exclude hyperparameter-dependent functions q)show ptabnew:update valid:0b from ptab where pnum>0 f | pnum pnames pvals valid ------------------------| ------------------------------------------------------ absenergy | 0 () () 1 abssumchange | 0 () () 1 count | 0 () () 1 countabovemean | 0 () () 1 countbelowmean | 0 () () 1 firstmax | 0 () () 1 firstmin | 0 () () 1 autocorr | 1 ,`lag ,0 1 2 3 4 5 6 7 8 9 0 binnedentropy | 1 ,`lag ,2 5 10 0 c3 | 1 ,`lag ,1 2 3 0 cidce | 1 ,`boolean ,01b 0 eratiobychunk | 1 ,`numsegments ,3 0 rangecount | 2 `minval`maxval -1 1 0 changequant | 3 `ql`qh`isabs (0.1 0.2;0.9 0.8;01b) 0 q)5#cfeatsnew:.ml.fresh.createfeatures[tab;`date;2_ cols tab;ptabnew] date | col1_absenergy col1_abssumchange col1_count col1_countabovemean col1_countb.. ----------| ---------------------------------------------------------------------------.. 2000.01.01| 1.33e+07 10100 30 13 17 .. 2000.01.02| 1.023e+07 11450 30 14 16 .. 2000.01.03| 7805000 9200 30 13 17 .. 2000.01.04| 8817500 9950 30 17 13 .. 2000.01.05| 7597500 7300 30 12 18 .. q)count 1_cols cfeatsnew / 74 columns now being created via a subset of initial functions 92
Modifications to the file
hyperparam.txt within the FRESH folder allows for fine tuning to the number and variety of calculations to be made. Users can create their own features by defining a function within the
.ml.fresh.feat namespace and, if necessary, providing relevant hyperparameters in .ml.fresh.params.
The operating principal of this function has changed relative to that in versions
0.1.x. In the previous version parameter #4 had been a dictionary denoting the functions to be applied to the table. This worked well for producing features from functions that only took the data as input (using
.ml.fresh.getsingleinputfeatures). To account for multi-parameter functions the structure outlined above has been used as it provides more versatility to function application.
Statistical significance tests can be applied to the derived features to determine how useful each feature is in predicting a target vector. The specific significance test applied, depends on the characteristics of the feature and target. The following table outlines the test applied in each case.
|feature type||target type||significance test|
Each test returns a p-value, which can then be passed to a selection procedure chosen by the user. The feature selection procedures available at present are as follows;
- The Benjamini-Hochberg-Yekutieli (BHY) procedure: determines if the feature meets a defined False Discovery Rate (FDR) level. The recommended input is 5% (0.05).
- K-best features: choose the K features which have the lowest p-values and thus have been determined to be the most important features to allow us to predict the target vector.
- Percentile based selection: set a percentile threshold for p-values below which features are selected.
Each of these procedures can be implemented by modifying parameter input to the following function;
Return statistically significant features based on defined selection procedure
tis the value side of a table of created features
tgtis a list of targets corresponding to the rows of table
fis a projection with example syntax
returns a list of features deemed statistically significant as deemed by the user defined procedure within parameter
q)tgt:value exec avg col2+.001*col2 by date from tab / combination of col avgs q)/ BHY procedure with a FDR level of 0.05 q)show sigBH:.ml.fresh.significantfeatures[value cfeats;tgt;.ml.fresh.benjhoch 0.05] `col2_mean`col2_sumval`col2_fftcoeff_maxcoeff_10_coeff_0_real`col2_fftcoeff_m.. q)/ Extract the top 20 best features q)show sigK:.ml.fresh.significantfeatures[value cfeats;tgt;.ml.fresh.ksigfeat 20] `mean_col2`sumval_col2`absenergy_col2`c3_1_col2`c3_2_col2`med_col2`quantile_0.. q)/ Extract the top 5th percentile of created features q)show sigP:.ml.fresh.significantfeatures[value cfeats;tgt;.ml.fresh.percentile 0.05] `col2_absenergy`col2_mean`col2_med`col2_skewness`col2_sumval`col2_c3_lag_1`co.. q)/ Check the count of each method to show differences in outputs q)count each (sigBH;sigK;sigP) 30 20 22
The v0.1.x input behaviour of
.ml.fresh.significantfeatures has been changed to accommodate an increased number of feature selection methods.