Hi @YSONG.1 ,
It depends -in general- on the specific application the customer is interested in.
I suggest the customer to look application_examples, where some use cases are presented (Activity recognition for wrist, Gym activity recognition, Yoga pose recognition).
You can find more info related to feature selection and other methods to avoid overfitting in our design tip document DT0139 (see the Feature selection, paragraph 2.2).
As a general statement, I would suggest to check the 1.3.13 paragraph of the AN5259 Application Note:
The selection of the features to be used for the Machine Learning Core configuration depends on the specific application. Considering that the use of too many features may lead to overfitting and too large decision trees, it is recommended to start first by selecting the four most common features:
- Mean
- Variance
- Energy
- Peak-to-peak
If the performance is not good with these features, and in order to improve the accuracy, other features can be considered to better separate the classes. Sometimes the basic features (mean, variance, energy, etc.) might not help in distinguishing the dominating frequency, so embedded digital filters can be enabled to select a specific region of frequency. Using the filtered signal, certain classes may be distinguished more precisely. For instance, if the user is walking, the typical signal is around 1-2 Hz, while if the user is jogging, the typical signal is around 2.5-4 Hz. The information contribution from a single feature can be evaluated by a measure of how much different classes are separated (from one another). This analysis can be done in a graphical way, by plotting 1D/2D graphs as described in the following examples.
-Eleon
