TS_SimSearch
.HCTSA_Empirical1000.mat
.TS_Init('INP_Empirical1000.mat');
to initialize, followed by compute commands involving TS_Compute
).hot_feature1
:boom:, recently published in Science (and not yet in the hctsa library), and attempt to determine whether it is completely new, or whether there are existing features that exhibit similar performance to it. Think first about the data context (described above), which allows you to understand the behavior of thousands of features on a diverse dataset with which to compare the behavior of our new feature, hot_feature1
. This example uses the Empirical1000
data context downloaded as HCTSA_Empirical1000.mat
from figshare.hot_feature1
, could be done directly (using TS_CalculateFeatureVector
), but in order to maintain the HCTSA structure, we instead produce a new HCTSA.mat
file containing just hot_feature
and the same time series. For example, to compare to the HCTSA_Empirical1000.mat
file hosted on figshare, you should use the same version of hctsa to enable a valid comparison to the same set of features.INP_hot_master.txt
containing the function call, that takes in a time series, x
:MyHotFeature.m
should be specified here. MyHotFeature.m
must also be in a form that outputs a structure (or a single real number, as explained here).MyHotFeature(x)
is hotFeature1
, which needs to be specified in another input text file, INP_hot_features.txt
, for example, as:hot
and science
.HCTSA_hot.mat
, containing information about the 1000 time series, and the new hot feature, hot_feature1
, which can then be computed as:TS_SimSearch
. We can find the ID assigned to our new hot_feature
in the merged HCTSA file as:my_hot_feature
in HCTSA_merged.mat
is 7703. Then we can use TS_SimSearch
to explore the relationship of our hot new feature to other features in the hctsa library (in terms of linear, Pearson, correlations):first_zero_ac
; see Interpreting Features for more info on how to interpret matching features):TS_SimSearch
provides another visualization of the context of this hot new feature (in this case there are so many highly correlated features, that the matrix doesn't reveal much subtle structure):first_zero_ac
), even across a highly diverse time-series dataset. However, if you have more luck and come up with a hot new feature that shows distinctive (and useful) performance, then it can be incorporated in the default set of features used by hctsa by adding the necessary master and feature definitions (i.e., the text in INP_hot_master.txt
and the text in INP_hot_features.txt
) to the library files (INP_mops.txt
and INP_ops.txt
in the Database directory of hctsa), as explained here. You might even celebrate your success by sharing your new feature with the community, by sending a Pull Request to the hctsa github repository!! :satisfied:HCTSA_Empirical1000.mat
on figshare :relaxed:fastdfa
algorithm from Max Little's website. This algorithm is already implemented in hctsa in the code SC_fastdfa.m
as the feature SC_fastdfa_exponent
. We can find the ID of this feature by finding the matching row in the Operations table (ID=750
):SC_fastdfa_exponent
, including some unexpected relationships with the stationarity estimate, StatAvl25
.TS_Plot2d
, e.g., as: