Enhancing Deep Consistent Graph Metric with Affinity and Alignment for Incremental Social Event Detection Using Cross-Layer Attention

Abstract

Existing methods of event detection from social media (i.e., X), for instance, KPGNN, FinEvent, and CLKD, use triplet loss for feature separation. Triplet loss suffers from two notable discrepancies in the latent space: (i) inconsistency in intra-event and inter-event distances, and (ii) an inability to ensure the closeness of messages from the same event across different mini-batches. The present paper proposes two novel loss functions to improve consistency in the latent space. The first loss function guarantees consistent intra-event and inter-event distances by increasing the affinity between intra-event points. On the other hand, the alignment loss enhances the cosine similarity between the feature space and label space, thereby aligning features of the same event class across diverse mini-batches. We provide theoretical justification that the proposed loss ensures discriminative features in the latent space, like CGML, without its costly pairwise or specialised batching. Adding to our loss function, we introduce a new attention module designed to effectively address heterogeneous relations without necessitating a separate optimisation objective. Through comprehensive experimentation on two publicly available datasets, we have shown an average improvement of $24.05\%$, $27.23\%$ and $123.69\%$ in NMI, AMI and ARI, respectively, over supervised SOTA event detection methods. Our method also shows improvements over SOTA unsupervised event detection methods across both datasets. These are supported by statistical significance tests. Generalizability of the proposed loss in general clustering problem in graph domain is shown through experiments.