Multi-label Learning with Semantic Correlations and Missing Labels

Abstract

We have witnessed the era of big data, when the massive amount of data can be downloaded/collected in different forms and diverse sources everyday. However, the correctness and completeness of the collected information cannot be guaranteed. In the supervised learning, the label information associated with samples is hard to validate in terms of its accuracy and completeness for its class dictionary, used methods and coding policy. Therefore, academia has seen a variety of approaches focusing on multi-class and multi-label problems. In the traditional multi-class learning, each sample is associated with a single label. In the novel multi-label learning, an object can belong to several labels simultaneously. Recently, multi-label learning has found numerous applications, in such as images, texts, music and bioinformatics. However, most of these methods are proposed under the assumption that a good format with complete labels is assigned to each instance. But in real-world applications, missing labels are inevitable. As a result, we handle the multi-label problem with considering missing labels.

Labels in multi-label learning are assumed to be correlated to each other. There are many research works that have demonstrated that semantic correlations can facilitate the multi-label classification. In this thesis, we introduce various methods that exploit different kinds of semantic correlations for multi-label learning with missing labels. First, the use of two-level semantic correlations is explored. An instance-wise semantic relational graph and a category-wise relational graph are exploited to perform label matrix imputation in the labeled space and label matrix prediction in the unlabeled space. Compared to the related graphs in other label recovery methods, these two graphs are extracted by considering semantic gaps and missing labels, respectively. Second, the low-rank and sparse semantic correlations are used to estimate missing labels when labels are locally correlated to each other in each topic of instances. The low-rank representation can reveal the membership of the label sets in a way that within-cluster affinities they are dense, while between-cluster affinities they are sparse. Compared to other methods, it is the first time that label sets are assumed to be in different topics. It means that label sets sharing the same cluster are strongly correlated to each other, while label sets of other clusters are loosely correlated to each other. Third, the semantic correlations based on topics are exploited to fill in missing labels. For all topics, semantic hierarchical correlations, which indicates the hyponymy and hypernymy relations between labels, are included to diversify the label dependency. Compared to other label imputation methods, we not only include the semantic hierarchy, but also learn the more reliable asymmetric label-wise semantic correlations based on the semantic hierarchy to facilitate the imputation. Instance-wise semantic correlations with semantic gaps oriented in each topic and two kinds of semantic correlations between topics are also extracted to infer the missing labels.

Compared to other state-of-the-art multi-label methods that deal with missing labels, the proposed methods mining three kinds of semantic correlations in this thesis are capable of exhibiting better performances, which are demonstrated by the intensive comparison studies.

Date of Award	30 Aug 2018
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Wai Shing Tommy CHOW (Supervisor)