💬 NLP

ValueGraph: Value-Signal Guided Graph Pre-training for Contextualized User Representation

The paper proposes ValueGraph, a graph pre-training framework that leverages inferred moral-value signals as soft constraints to enhance contextualized user representations, demonstrating consistent performance improvements in stance detection and bot detection tasks over existing text-based, graph-based, and LLM baselines.

Original authors: Yitong Han, Wei Gao, Yi Zhao, Prasanta Bhattacharya, Fengzhu Zeng, Mohammad Amanlou

Published 2026-09-02

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Yitong Han, Wei Gao, Yi Zhao, Prasanta Bhattacharya, Fengzhu Zeng, Mohammad Amanlou

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the vast, noisy landscape of social media, understanding why people act the way they do is a central challenge for anyone trying to make sense of online life. For years, computer scientists have tried to build digital profiles of users by looking at what they say and who they talk to. If two people post similar words or reply to the same threads, traditional systems assume they are alike. However, this approach often misses a crucial layer of human behavior: the underlying moral values that drive those interactions. People might engage with the same controversial news story for entirely different reasons—one out of genuine concern, another out of anger, and a third out of a desire to spread confusion. These hidden motivations shape how users behave, yet standard computer models often treat them as invisible. To understand a user truly, researchers need a way to capture not just the surface of their activity, but the moral compass that guides it.

A team of researchers at Singapore Management University and A*STAR has developed a new method called ValueGraph to solve this problem. Instead of trying to guess a person's true inner beliefs, which is impossible to know for certain, the system uses a tool to infer a "value signal" from their public posts. This signal is based on a well-known psychological framework called Moral Foundations Theory, which breaks down moral reasoning into ten distinct dimensions, such as care, fairness, loyalty, and authority. The researchers trained a computer program to read thousands of social media posts and assign a score to each one based on these ten moral themes. By averaging these scores across all the posts a single user has written, the system creates a rough, ten-point profile of that user's moral framing. The researchers are careful to note that this is not a perfect psychological diagnosis; it is a noisy, aggregate signal that captures the general flavor of a user's moral language rather than their exact soul.

The core innovation of ValueGraph lies in how it uses these inferred signals to teach a computer to recognize different types of users. The system operates in two main steps. First, it builds a map of how people interact, treating every post and reply as a connection in a giant web. It learns to understand the structure of these conversations and the meaning of the text within them, creating a basic digital fingerprint for every user. In the second step, it brings in the moral value signals. The system looks at the inferred moral profiles of different users and groups those with similar value signals together, while pushing those with very different signals apart. It does this by treating the moral signals as a set of soft rules or gentle nudges during the learning process, rather than as absolute facts. This allows the computer to learn a representation of a user that combines what they say, who they talk to, and the moral tone of their expression.

The researchers tested this new approach on two difficult tasks: determining a user's stance on a political issue and detecting whether a user is a bot, or an automated account designed to mimic a human. In the first test, the system had to figure out if a user supported, opposed, or was neutral toward a specific topic. In the second, it had to distinguish between real humans and automated bots. The results showed that ValueGraph consistently outperformed existing methods. On the stance detection task, it achieved an accuracy of 63 percent on one dataset and 77 percent on another, beating strong competitors that relied only on text or only on network structure. For the bot detection task, the system improved the ability to spot automated accounts, particularly when combined with other data. In one specific configuration, the system's ability to correctly identify bots improved significantly, with the "recall" rate—the ability to find the bots when they are there—jumping from 62.5 percent to 81.0 percent when using the new method.

What makes these results particularly interesting is that the system does not need to be told which users are bots or which stance is correct during its initial training. It learns these patterns by observing the moral signals and the structure of the conversations on its own. The researchers found that the system worked best when it combined the structural learning with the moral value guidance. If they removed the moral signals, the system's performance dropped. If they used only the moral signals without the conversation structure, it also failed to reach the same level of accuracy. This suggests that the moral value signal acts as a useful guide, helping the computer see patterns that are invisible when looking at text or connections alone. For instance, in visualizing the data, the researchers saw that the new method created much clearer separation between human users and bots than previous methods did. The bots tended to cluster together with repetitive, polarized moral language, while humans showed more varied and context-sensitive expressions.

The team emphasizes that their method does not claim to know a user's true psychological values. The moral scores are derived from text and are treated as imperfect, noisy data. The system uses these imperfections as a learning tool, finding the relative differences between users rather than trying to pin down an absolute truth. This approach avoids the pitfalls of trying to label people with definitive psychological profiles, which can be ethically fraught and technically unreliable. Instead, it uses the moral framing of their public discourse as a practical tool to build better, more nuanced digital representations. The success of ValueGraph suggests that incorporating these theory-informed value signals provides a powerful way to improve how computers understand human behavior online, offering a more robust foundation for tasks ranging from content moderation to personalized recommendations. By acknowledging that users are shaped not just by what they say, but by the moral lens through which they say it, the researchers have opened a new path for making sense of the complex social world of the internet.

Technical Summary: ValueGraph

Problem Definition
Existing user representation methods in social media analytics primarily rely on observable behaviors, such as textual content, visual media, or network interaction structures. These approaches often assume that users with similar interaction patterns or content exposure share similar underlying motivations. However, the authors argue that this surface-level similarity is brittle; users may engage with the same events due to shared exposure or controversy while holding fundamentally different attitudes and values. This limitation can lead to spurious behavioral associations in downstream tasks (e.g., recommendation systems reinforcing polarization) because conventional models fail to capture the latent moral and value-related factors driving user behavior. Furthermore, existing graph pre-training methods rarely incorporate value signals or motivational dimensions as explicit inductive biases.

Methodology: ValueGraph Framework
The paper proposes ValueGraph, a graph pre-training framework designed to learn contextualized user representations by integrating automatically inferred moral-value signals as noisy auxiliary supervision. The framework operates on post-reply graphs and proceeds in two stages:

Stage 1: Foundation GNN Pre-training
- Input: A constructed post-reply graph where nodes represent posts and edges represent reply interactions.
- Encoding: Posts are encoded using ModernBERT to handle noisy social media text.
- Objective: The model undergoes unsupervised pre-training using a masked autoencoding paradigm (based on GraphMAE2). Node features are masked, and the model reconstructs them. This stage learns noise-resilient semantic and structural representations of posts, which are then aggregated into initial user embeddings.
Stage 2: Value-Guided Pre-training
- Value Signal Derivation: The framework employs MoralBERT (a fine-tuned language model based on Moral Foundations Theory) to infer moral value scores for each post across 10 dimensions (e.g., care, harm, fairness, loyalty). These post-level scores are aggregated to create a 10-dimensional value signal vector for each user. Crucially, these vectors are treated as noisy auxiliary signals rather than gold-standard psychological labels.
- Contrastive Pair Construction: Users are paired based on the similarity of their inferred value vectors. A symmetric percentile-based strategy defines positive pairs (high value similarity) and negative pairs (low value similarity).
- Hierarchical Objective: The framework optimizes user embeddings using a combination of two loss functions:
  - User-level Contrastive Loss ( $\mathcal{L}_{user}$ ): An InfoNCE loss that pulls embeddings of users with similar value profiles together and pushes dissimilar users apart.
  - Clustering Loss ( $\mathcal{L}_{cls}$ ): A global regularizer that periodically performs K-means clustering on user embeddings. It includes a compactness term (encouraging users to stay close to their cluster centroid) and a margin term (enforcing separation between distinct cluster centroids).
- Total Loss: The final objective is a weighted sum: $\mathcal{L} = \mathcal{L}_{user} + \lambda \mathcal{L}_{cls}$ . This design ensures that the model learns representations that satisfy both local pairwise value consistency and global structural coherence, acting as mutual regularizers to filter out noise.

Key Contributions

ValueGraph Framework: A novel graph pre-training approach that leverages inferred moral-value signals as noisy auxiliary supervision to enhance contextualized user representation learning.
Hierarchical Objective Design: A training strategy combining masked graph autoencoding with value-guided contrastive learning and clustering to jointly capture semantic, structural, and value-relevant information.
Mechanism-Level Analysis: Theoretical characterization demonstrating how the objectives preserve value-signal similarity ordering and promote compact, separated user groups in the embedding space.
Empirical Validation: Comprehensive experiments showing that ValueGraph outperforms strong baselines in downstream tasks without requiring explicit value labels for the downstream tasks themselves.

Experimental Results
The authors evaluated ValueGraph on two downstream tasks: Stance Detection and Twitter Bot Detection.

Stance Detection: Evaluated on MT_CSD and RumourEval19 benchmarks.
- ValueGraph achieved 63% accuracy and 58% macro-F1 on MT_CSD, outperforming the strongest pre-trained baseline (GraphMAE2) by 5% in accuracy and 5.5% in macro-F1.
- On RumourEval19, it achieved 77% accuracy and 71% macro-F1, representing a 20.3% relative improvement in macro-F1 over the best PLM baseline (ModernBERT) and a 57.8% improvement over the best graph-based baseline (GTN).
- Performance was consistent across graph-based, text-based, and LLM baselines (including GPT-5.4).
Twitter Bot Detection: Evaluated on the TwiBot-22 dataset using various baseline architectures (RoBERTa, BotGCN, BotGAT, BotRGCN).
- Replacing baseline text or graph embeddings with ValueGraph embeddings consistently improved performance.
- For example, in a text-only setting with RoBERTa, macro-F1 increased from 54.8% to 64.3%, and recall improved from 62.5% to 81.0%.
- In multimodal settings (Text + Network + Features), ValueGraph embeddings still provided statistically significant gains, particularly in macro-F1 and Matthews Correlation Coefficient (MCC), suggesting the value signals capture complementary cues for distinguishing bots from humans.
Ablation Studies: Removing either the contrastive loss ( $\mathcal{L}_{user}$ ) or the clustering loss ( $\mathcal{L}_{cls}$ ) resulted in inconsistent performance, often improving accuracy at the cost of macro-F1. The full model, integrating both, yielded the best results, validating the necessity of the dual-objective design.
Visualization: t-SNE visualizations showed that ValueGraph embeddings produced clearer separation between bot and human users compared to text-only or standard graph-based encoders.

Significance and Claims
The paper claims that Value-signal guidance serves as a useful inductive bias for socially informed user modeling. By treating inferred moral values as soft constraints rather than ground truth, ValueGraph learns representations that capture not just what users say or who they interact with, but the underlying value-driven tendencies of their behavior.

The authors emphasize that their approach does not attempt to predict users' "true" psychological values. Instead, it utilizes theory-informed (Moral Foundations Theory), automatically inferred signals to regularize the representation learning process. The results suggest that even noisy, aggregate value signals can significantly improve the robustness and generalizability of user embeddings for tasks like stance detection and bot identification, addressing the brittleness of models that rely solely on interaction similarity.

Technical Summary: ValueGraph

More like this