How political views change on social media: a new TSN4PI model is presented
A new framework, TSN4PI, combines large language models and temporal graph neural networks to identify users' political views and predict their changes. The authors also released two large datasets and validated the approach on data from X and Truth Social.

Why traditional approaches fall short
Social media has become the main arena for political discussions. Posts, likes, and reposts reflect how people's worldviews shift, but researchers struggle to keep track of this process. The main obstacles have long been known:
- Politically significant messages get lost in the flood of entertainment and personal content.
- Manual annotation for training algorithms is expensive and suffers from subjectivity.
- Data from different platforms is structured differently, so a model trained on one social network does not work on another.
- The hardest part is predicting how a person's views will change, since they are influenced by news, their environment, and the behavior of other users.
Traditional machine learning methods often fail: they either miss the relevant signals or merely describe current preferences.
How TSN4PI works
TSN4PI is a new system for tracking the evolution of political views in social media. Its architecture is split into two parts, each tackling its own task.
Ideological filtering
The first module (PIDN) is responsible for separating political content from irrelevant material. It leverages large language models and style-transfer techniques, which allows it to process noisy data from different sources without additional annotation. In simple terms, the module learns to understand what a post is actually about and not get distracted by entertainment content.
Predicting shifts
The second module (PIPN) looks to the future. Using temporal graph neural networks, it analyzes a user's connections to other people and posts and estimates which direction their stance will move: whether the existing ideology will strengthen, weaken, or shift entirely to the opposite.

What testing showed
The researchers collected two large datasets related to the X and Truth Social platforms and made them available for non-commercial research. Evaluation showed that TSN4PI successfully handles its tasks. The system not only detects traces of political beliefs but also captures changes that occur over time.
Of particular interest are the differences between platforms. On X and Truth Social, audience polarization manifests differently, and the model makes it possible to see these nuances. This is empirical confirmation that political views online are not static: they drift under the influence of the information environment.

Why this matters
The ability to accurately track the evolution of ideologies matters not only for theoretical research. It helps predict public sentiment, identify manipulative patterns, and make recommendation algorithms more neutral. Thanks to the open datasets, other teams can test their own approaches on the same data, making results easier to compare.
The model is not perfect — human political behavior is too complex to fit into a single formula. But such a unified framework sets a benchmark to build on in future work. The study has already been accepted for publication in the reputable journal ACM Transactions on Intelligent Systems and Technology.
Frequently asked questions
Related materials
All materials
Canonical Polynomial Invariants for Diagnosing RL Simulator Failures: Where Are They Actually Useful?
The study examines whether incorporating physical laws into learnable dynamics models is justified: invariants extracted from trajectories and reduced to canonical form using Gröbner bases yield modest forecasting gains, but prove powerful for diagnostics, enabling the localization of violated constraints and identification of specific erroneous parameters.
10 November 2026
AURA-Eval: How to Check Whether an LLM Agent Recognized a Risk and Changed Its Next Step
The framework evaluates LLM agents’ behavior not by their final verdict, but by how they respond to risk at key decision points: tool-call trajectories are interrupted, after which the model must continue reasoning and acting, while a panel of LLM judges assesses risk detection, the nature of the step, and its safety. Across 1,249 tasks drawn from 157 R-Judge trajectories, unsafe actions become more frequent among 20 models when no safe path is available, and removing oversight before execution turns 52.3% of initially safe cases into unsafe ones.
10 November 2026
Chern–Simons Reservoir Memory: Autonomous Connectivity Dynamics and Density Transport
The article explores how a reservoir preserves the history of input signals both in its density distribution and in its geometric degrees of freedom. A comparison of four mechanisms shows that current creates a separate connected component that alters the subsequent density response, while accounting for all measured features yields a small improvement in decoding pulse order.
10 November 2026