Deep Forest with Sparse Topological Feature Extraction and Hash Mapping for Brain Network Classification

As a novel non-neural network style deep learning method, the deep forest can perform effective feature learning without relying on a large amount of training data, thus brings us some opportunities to accurately classify brain networks (BNs) on limited fMRI data. Currently, preliminary attempts to use deep forest to classify BNs are already emerging. However, these studies simply adopted the sliding windows to scan the inputted BNs and failed to consider the inherent sparsity of BNs, which makes them susceptible to those redundant edges in BNs with little weight. In this paper, we propose a deep forest framework with sparse topological feature extraction and hash mapping (DF-STFEHM) for BN classification. Specifically, we first design an extremely random forest guided by a weighted random walk (ERF-WRW) to extract sparse topological features from BNs, where the random walk strategy is used to capture their topological structures and the weighted strategy is used to reduce the influence of redundant edges with little weight. Then, we map these sparse topological features into a compact hashing space by a kernel hashing, which can better preserve topological similarities of brain networks in the hashing space. Finally, the obtained hash codes are fed into the casForest to perform deeper feature learning and classification. Experimental results on ABIDE I and ADHD-200 datasets show that the DF-STFEHM outperforms several state-of-the-art methods on classification performance and accurately identifies abnormal brain regions.