Polarization Detection on Social Networks: dual contrastive objectives for Self-supervision

TL;DR

This paper addresses polarization detection on social networks by introducing DocTra, a unified self-supervised framework with dual contrastive objectives: an interaction-level objective that contrasts positive and negatives interactions (including polarization-induced silence) and a feature-level objective that decouples polarized from invariant features. It provides an efficient solver, supports semi-supervised and prompt-tuning supervision, and proposes a unified polarization index to quantify polarization while normalizing background engagement and mitigating outliers. Empirical results on seven public datasets show significant improvements over eight baselines, demonstrating robustness to varying edge types, signs, and noise. The work advances polarization analysis by delivering a generalizable, self-supervised approach with practical utility for clustering, classification, and cross-dataset comparison.

Abstract

Echo chambers and online discourses have become prevalent social phenomena where communities engage in dramatic intra-group confirmations and inter-group hostility. Polarization detection is a rising research topic for detecting and identifying such polarized groups. Previous works on polarization detection primarily focus on hand-crafted features derived from dataset-specific characteristics and prior knowledge, which fail to generalize to other datasets. This paper proposes a unified self-supervised polarization detection framework, outperforming previous methods in unsupervised and semi-supervised polarization detection tasks on various publicly available datasets. Our framework utilizes a dual contrastive objective (DocTra): (1) interaction-level: to contrast between node interactions to extract critical features on interaction patterns, and (2) feature-level: to contrast extracted polarized and invariant features to encourage feature decoupling. Our experiments extensively evaluate our methods again 7 baselines on 7 public datasets, demonstrating significant performance improvements.

Figures (5)

• Figure 1: Toy example of a polarization detection task: The input consists of up to 3 types of edges (user-to-user, user-to-post, and post-to-post) of up to 2 types of signs (positive and negative).
• Figure 2: Dual contrastive objectives for self-supervised polarization detection: (a). contrast between positive interactions (what the user interacts with) and sampled 'negative' interactions (what the user does not interact with). The red dashed lines present the possible sampled 'negative' interactions. The key challenge is to eliminate false negatives and ineffective negative pairs. (b). contrast between polarized and invariant features.
• Figure 3: Iterative framework of DocTra: (a). obtain class assignments from current embeddings; (b). obtain decoupled features; (c). given the anchor node (red), sample positive interactions (green line) and negative interactions (red dash line) by augmenting the anchor node and solving eq.(1); (d). performing contrastive learning on both objectives to update the embeddings.
• Figure 4: Prompt-tuning framework: the triangle nodes are the learnable prompt nodes added to the input graph. The dashed lines are the induced edges derived from Connect(,).
• Figure 5: Polarization classification with semi-supervision