XML-CNN

A Deep Learning model optimizing Convolutional Neural Networks for Extreme Multi-label Text Classification.

XML-CNN addresses the challenge of tagging documents with thousands of potential labels simultaneously. Developed by researchers at Carnegie Mellon University and IBM, the architecture improves upon traditional CNNs by utilizing a bottleneck layer to compress high-dimensional features and a dynamic pooling strategy to capture global document semantics. In benchmarks like the EUR-Lex and Amazon datasets, XML-CNN outperforms conventional models (such as FastText) by effectively handling label sparsity and capturing local correlations between words. It is a go-to framework for large-scale categorization tasks where precision across massive label sets is critical.

https://arxiv.org/abs/1709.07155