Relation Guided Message Passing for Multi-label Classification
As the name implies, multi-label classification entails selecting the correct subset of tags for each instance. Its main difference from multi-class learning is that the class values are not mutually exclusive in multi-label learning and usually no prior dependencies between the labels is provided explicitly. In general, the existing literature treats the relationship between labels as co-occurrences in an undirected manner. Though, there are usually multiple types of dependencies between labels and their strength s, they are not independent from the direction of the specified edge. For instance, the “ship” and “sea” labels have an obvious dependency, but the presence of the former implies the latter much more strongly than vice versa. In this project, we introduce relational graph neural networks to model label dependencies. We consider two types of statistical relationships; pulling and pushing.
The input features pass through embedding initialization and self attention layers of MHA (Multi-Head Attention) and a PFF (Position-wise Feed Forward network) to initialize input embeddings. These are used as input label node embeddings. The Label Dependency Module (LDM) layers update the embeddings by relation-based composition, direction-specific neighbourhood aggregation and relation-conditioned activation by pulling and pushing relation sets. The label predictions are obtained by reshaping the embeddings, and then passing them through a final feedforward layer and a sigmoid to normalize.
The experiments show that the integration of the multi-relation module with pulling and pushing relations allows the architecture to emulate the advantage of making predictions sequentially in RNN based architectures, while enjoying the flexibility and computational efficiency of an MPNN.
The first sample has ground truth labels 'radio' and 'music'. The induced label sub-graph shows that the relevant labels pull each other, push the label `video' at first hob and push 'internet' at second hob. Based on the predicted label likelihoods, it can be observed that the integration of LDM module reinforce the belief on relevance of 'music' and redundancy of 'video' and 'internet'. The second example is tagged only with the label 'shipyard' which pushes 'search' at first hob and 'searchengine' at second hob, accordingly the LDM module amplify the difference between the likelihoods of relevant and irrelevant labels. Note that the induced label sub-graphs are obtained by lowest optimal hard prediction threshold observed during the experiments of dataset, therefore the changes in the predicted likelihoods are critical to make a decision towards capturing the full label set.
Team
Publications
- M. Ozmen and M. Coates. Relation Guided Message Passing for Multi-label Text Classification. In submission process, 2023.
- M. Ozmen, H. Zhang, P. Wang, and M. Coates. Multi-relation message passing for multi-label text classification. In Proc. IEEE Int. Conf. Acoustics, Speech and Signal Processing (ICASSP), 2022.