Extreme Multi-label Classification from Aggregated Labels

Extreme multi-label classification (XMC) is the problem of finding the relevant labels for an input from a very large universe of possible labels. XMC has wide applications in machine learning including product categorization [AGPV13, YJKD14], webpage annotation [PKB 15] and hash-tag suggestion [DWP 15], where both the sample size and the label size are extremely large. Recently, many XMC methods have been proposed with new benchmark results on standard datasets [PKG 18, GMW 19, JBCV19]. XMC problem, as well as many other modern machine learning problems, often require a large amount of data. As the size of the data grows, the annotation of the data becomes less accurate, and large-scale data annotation with high quality becomes growingly expensive. As a result, modern machine learning applications need to deal with certain types of weak supervision, including partial but noisy labeling and active labeling.