simple correct&smooth (cands)

Simple implementation of correct and smooth (C&S) in pytorch. cands is intended to be very quick for practictioners to pick up and run as a postprocessing step on your current classifiers to add signal from existing (but unused) graph structure in your data. While lots of data contains some form of graph structure, it is rarely used in practice due to the challenges of setting up GNNs or extracting manual graph features.

What tasks work well with C&S

C&S works for transductive node classification tasks, which means that the graph itself remains static and accessible during both traing and validation. After you have generated predictions, C&S will "smooth" them using the neighboring training node features and labels. We've seen up to 20% increase in validation performance when running C&S over neural network-based predictions on the features alone.

Installation

Install the latest version with:

pip install https://github.com/andrewk1/correctandsmooth/archive/refs/tags/v0.0.2.zip

Install from source

git clone https://github.com/andrewk1/correctandsmooth.git && cd correctandsmooth
pip install .

cands has been tested on torch>=1.9.

Usage

C&S is a single postprocessing step applied to your predictions, we expose it via a function

from cands import correct_and_smooth

yhat_cands = correct_and_smooth(y, yhat, edge_index, val_split_idxs)

Currently, C&S needs four arguments:

y: Your current labels in matrix form (N, D) where N is your dataset length and D is the number of classes.
yhat: Predictions of shape (N, D)
edge_index: Tensor of shape (2, E). Columns contain indices 0<=i<N of nodes with shared edges. E is the total edge count
val_split_idxs: Indices of validation or unlabeled nodes (nodes not in the training set)\