Tree structured neural networks¶

Tree structured neural networks extend deep learning to hierarchical data organized as rooted trees or directed acyclic graphs. Common applications in misinformation research include modeling social media conversation threads, propagation cascades, and argument structures.

Motivation¶

Traditional sequential neural networks (RNNs, LSTMs) assume linear dependencies. Tree structures capture: - Conversation hierarchies: Replies to replies form a tree with the root post at the top - Propagation patterns: Information spreads from a source through retweets/shares forming tree structures - Argument structures: Counterarguments, evidence, and premises form discourse trees - Document hierarchy: Paragraphs → sentences → words form a tree

Tree-aware architectures learn to aggregate information from children nodes efficiently, capturing patterns that depend on tree topology.

Architectures¶

Recursive Neural Networks (RvNNs)¶

Process trees bottom-up: leaf node representations are encoded, then parent nodes recursively combine child representations. Applied to rumor detection by Ma et al. (2018) who show propagation tree topology carries veracity signals.

Tree LSTMs¶

Extend LSTM cells to tree structures with child aggregation mechanisms. Each node processes text input plus aggregated child hidden states: - Child Sum: Average or sum hidden states from all children - Child Max-Pooling: Take element-wise maximum across children - Child Convolve: Apply convolution operation to learn patterns in child features

Kumar & Carley (2019) propose convolutional aggregation units for learning source-reply contrasts in conversation trees for stance and rumor veracity classification.

Binarized Tree Models¶

Convert n-ary trees (variable number of children per node) to binary trees (maximum two children) via virtual parent nodes. Benefits: - Enables standard binary tree operations - Can force model to learn pairwise relationships - Tradeoff: adds artificial nodes that complicate class labeling in some tasks

Applications in misinformation research¶

Rumor veracity classification: Tree topology contains signals about truth value; top-down and bottom-up message passing captures how claims are disputed or supported
Stance classification: Reply-to-root relationships in conversation threads signal whether replies support, deny, or question the root claim
Propagation analysis: Tree structure of retweet cascades reveals user behavior (early amplifiers, late echoes) and content virality
Argument mining: Document structure trees (discourse parsing) reveal argumentative relationships needed for fact-checking

Key papers in this wiki¶

Kumar & Carley (2019) — Tree LSTMs with Convolution Units to Predict Stance and Rumor Veracity in Social Media Conversations — proposes Tree LSTM and Binarized Constituency Tree (BCTree) LSTM with convolution units; achieves state-of-the-art on PHEME dataset for stance (0.520 F1) and veracity (0.379 F1) classification; demonstrates convolutional child aggregation outperforms sum/max-pooling.
Zubiaga et al. (2016) — Stance Classification in Rumours as a Sequential Task Exploiting the Tree Structure of Social Media Conversations — linear and tree CRF models; demonstrates that tree-structured modeling improves stance classification over linear sequential models.
Ma et al. (2018) — Rumor Detection on Twitter with Tree-structured Recursive Neural Networks — applies RvNNs to propagation trees; achieves 72.3%/73.7% accuracy on Twitter15/16; shows tree topology captures veracity signals.
Ma et al. (2017) — Detect Rumors in Microblog Posts Using Propagation Structure via Kernel Learning — Propagation Tree Kernel matching similarity between propagation trees; 75% accuracy with superior early detection.

Graph Neural Networks (GNNs): Generalize trees to arbitrary graphs; used for social network-based rumor detection
Recursive neural networks: Similar motivation of processing nested structures
Attention mechanisms: Alternative to tree structure for learning which nodes matter