GCN - BASED

FEATURE
SELECTION
 Method
Algo iteratively finds a set of optimal features.
The iterative algorithm which has five major components:
• feature initialization,
• graph construction,
• neural network,
• multiple dropouts,
• gradient computation

aims to iteratively find a set of optimal features which gives rise to the
greatest decreases in the optimization loss.
Features

For feature initialization, given a feature matrix.

we first introduce a bias feature.
The total number of features now is p + 1.
the bias feature serves as the initial selected feature to start the feature
selection process.
Graph construction

Considering each sample as a node, we connect two nodes if their

cosine similarity score is larger than a threshold δ which is a
hyperparameter.

The resulting similarity graph captures the latent interactions between

samples and will be used in the graph convolutional network (GCN)
layer. The similarity graph is different at each iteration.
GCN

GCNs—GraphSAGE leverages node feature information to efficiently

generate embeddings by sampling and aggregating features from a
node’s local neighborhood.
Results

Reported the average test AUROC (over 20 times train test splits) with
respect to the number of selected features from 1 to 10.
Inferences
reported that it might fail on data with highly nonlinear relations
between features and labels.It is computationally inefficient.

gradient estimation
The gradient of the loss function, in simple terms, tells you how much
changing a particular input to your model will change the model's
overall error (loss). It's like a map showing you which directions to
move in to quickly decrease the error and improve your model's
performance.
 Other scope

Feature Graphs

Modeling the interactions

between features of a data
instance as a homogeneous
graph, in which nodes are
features and edges
are their correlation.
 Other scope

GAT

Unlike conventional GNNs which aggregate features from all

neighbors equally, GATs employ an attention mechanism. This
allows them to focus on the most informative neighbors for each
node, effectively learning feature importance implicitly. This leads
to better discrimination between relevant and irrelevant features.