Multi-label Learning with Missing Labels

Abstract

In multi-label learning, each sample can be assigned to multiple class labels simultaneously. In this work, we focus on the problem of multi-label learning with missing labels (MLML), where instead of assuming a complete label assignment is provided for each sample, only partial labels are assigned with values, while the rest are missing or not provided. The positive (presence), negative (absence) and missing labels are explicitly distinguished in MLML. We formulate MLML as a transductive learning problem, where the goal is to recover the full label assignment for each sample by enforcing consistency with available label assignments and smoothness of label assignments. Along with an exact solution, we also provide an effective and efficient approximated solution. Our method shows much better performance than several state-of-the-art methods on several benchmark data sets.

ML-MG: Multi-label Learning with Missing Labels Using a Mixed Graph

“…Learning from such partially labeled instances is referred to as the multi-label learning with missing labels (MLML) problem [37,43]. As labels are usually related by semantic meanings or co-occurrences, the key to filling and learning from missing labels is a good model to represent label dependency.…”

“…For example, the label dependency between a pair of labels, such as instance similarity and class cooccurrence can be represented using such a graph (see green and blue edges in Figure 1). However, as stated in [37,39], class co-occurrence that is derived from training labels can be inaccurate and even detrimental when many missing labels exist. Li et al [23] propose to alleviate this limitation by using an auxiliary source (such as Wikipedia) to estimate co-occurrence relations.…”

“…Firstly, motivated by [37,39], a discrete objective function is formulated to simultaneously encourage consistency between predicted and ground truth labels and encode traditional label dependencies (instance similarity and class co-occurrence). Whereas, semantic hierarchy constraints are incorporated as hard linear constraints in the matrix optimization.…”

“…Third, missing labels can be treated as latent variables in probabilistic models, including the model based on Bayesian network [22,31] and conditional restricted Boltzmann machines (CRBM) model. Last, Wu et al [37] define three label states, including positive labels +1, negative labels −1 and missing labels 0, to avoid the label bias. However, the two solutions proposed in [37] involves matrix inversion, which limits the scalability to handle larger datasets.…”

“…Last, Wu et al [37] define three label states, including positive labels +1, negative labels −1 and missing labels 0, to avoid the label bias. However, the two solutions proposed in [37] involves matrix inversion, which limits the scalability to handle larger datasets. Wu et al [39] propose an inductive model based on the framework of regularized logistic regression.…”