FeaSel-Net 0.0.3

A Keras callback package for iteratively selecting the most influential input nodes during training.

SpecNet

SpecNet is a python package that enables feature selection algorithms embedded in a neural network architecture. It combines a leave-one-out cross-validation (LOOCV) type of feature selection algorithm with recursive pruning of the input nodes, such that only the most relevant nodes with the richest information are kept for the subsequent optimization task. The recursive pruning is undertaken by employing a FeatureSelection callback at certain points of the optimization process. The precise procedure is explained in Sequence of Events. Originally developed for serving the task of finding biomarkers in biological tissues, the algorithm is generically coded such that it is able to select features for all kinds of classification tasks.

The package is an extension for the keras and tensorflow libraries. Please see the links for further information on their software packages and to get a grasp of neural networks in general and the constructs used for SpecNet.

Installation

It is best at the moment to install this as an external package with pip. This can be done by cloning the repository with the following commands:

git clone https://github.tik.uni-stuttgart.de/FelixFischer/FeaSel-Net.git feasel-net
cd feasel-net
pip install -e .

Sequence of Events

1. Initiallizing Neural Network The first step of the algorithm can be thought of a simple optimization problem initiallized with the inputs and a binary mask for those inputs with only ones as its entries. This behaviour is induced by using a newly created layer type called LinearPass.

   

2. Training until trigger conditions are met The neural network optimizes the classification results until one of the following options happen:

   • the training (or validation) loss value is beneath a certain threshold
   • the training (or validation) accuracy value is above a certain threshold Then - for the sake of consistency - it will count how many times in a row the conditions are met. If this happens for multiple epochs, the actual pruning event will start that consists of estimating the importance and eliminating uninformative features.

3. Importance estimation As soon as the callback is triggered, the evaluation of the

   

Release Information

0.0.1 - Initial Release

• callback FeatureSelection
  • trigger parameters: delta epochs, thresholds, ...
  • different metrics for triggering
  • etc.
• layer LinearPass

ToDos

Until now, only dense layered architectures are supported. The plan is to also include convolutional layers.

[x] DenseLayer support

[x] accuarcy and loss based evaluation

[ ] ConvLayer support

[ ] intermediate layers shall be supported

[ ] paper on algorithm