PaperlayerView exact revision on arXiv

Source-linked AI summary

Against Political Polarization: A Unified Framework for Tracing Evolving Political Ideologies on Social Media

Yijie Xu, Chao Wang, Hui Xiong

arXiv:2608.17987v1cs.SIcs.AIcs.CLcs.LG

TL;DR

Social media makes ideological expression and polarization increasingly important to study, but research faces limited data access and noisy content. TSN4PI combines ideology detection with temporal graph modeling, finding that the most ideologically volatile users tend to converge toward the center.

  • Problem

    Research on social-media polarization and ideologies faces scarce suitable datasets and abundant content, while echo-chamber significance and mechanisms remain debated.

  • Method

    TSN4PI combines LLM-based style transfer and unsupervised domain adaptation for ideology detection with temporal graph neural networks for predicting users’ ideological evolution.

  • Results

    The most ideologically volatile users tend to converge toward the center, contrary to widely held assumptions about increasing online polarization.

  • Takeaways & Limitations

    The analysis offers empirical evidence for a more nuanced account of online ideological evolution than increasing polarization alone.

  • Takeaways & Limitations

    LLM-based style transfer may introduce ideology drift, lexical homogenization, or source leakage across ideology groups and platforms.

Abstract

from arXiv · show

The rapid growth of social media has greatly influenced political discourse, highlighting the need to understand individual political ideologies and their temporal dynamics. This task faces challenges such as data scarcity, abundant non-political content, costly and bias-prone manual annotation, and difficulty in modeling future ideological inclinations. To address these issues, we propose TSN4PI, a unified framework for tracking the evolution of political ideologies on social media. It includes two core modules. The PIDN uses large language models with style transfer and unsupervised domain adaptation to enable robust ideology detection and filter irrelevant content from noisy, cross-domain data. The PIPN employs temporal graph neural networks to predict future ideological shifts, enabling comprehensive analysis of ideology presence, intensity, and evolution. We release two large-scale datasets for noncommercial research use to facilitate further work. Extensive case studies on multiple platforms (X and Truth Social) validate the effectiveness of TSN4PI and provide empirical insights into political polarization and the evolution of online ideologies. Our findings offer a nuanced perspective, advancing both methodological development and empirical understanding in this field.

1 Introduction … Preliminary Data Analysis.

The paper frames political ideology as an evolving, broad belief system shaped by social-media discourse and introduces TSN4PI to detect, quantify, and predict ideological dynamics across noisy platforms. It combines cross-domain ideology detection, temporal prediction, multi-platform datasets, and preliminary linguistic analysis.

  • 1 Introduction: Social media enables ideological expression and dissemination at scale while algorithmic curation and social clustering can reinforce echo chambers and political polarization.
  • 2 Related Work: Existing research spans ideological clustering and event-based polarization studies, but the significance and mechanisms of echo chambers remain debated.
  • 1 Introduction: TSN4PI unifies PIDN for ideology detection and intensity estimation with PIPN for modeling and predicting users’ ideological evolution.PIDN combines LLM-based style transfer with unsupervised domain adaptation, while PIPN leverages temporal graph neural networks.
  • 2 Related Work: Political ideology is broader and more stable than target-specific, transient stance, so this study models ideology across topics and over time.
  • 3 Data: The study uses AllSides news-source labels plus user-centric posts from X and Truth Social for training, evaluation, adaptation, and cross-platform analysis.The AllSides dataset supports transferring ideological signals across domains, while the social-media datasets represent distinct online communities.
  • Preliminary Data Analysis.: The AllSides news dataset has a relatively balanced Left–Right distribution, supporting class balance, while its 20 largest outlets span all five ideological categories.The analysis examined article and outlet proportions across five political categories and outlet concentration.
  • Preliminary Data Analysis.: Trump is the most frequent term in both Left (96,727) and Right (60,993) articles, indicating shared attention with potentially divergent framing.The frequency analysis removed stopwords and compared the top 10 content words for each political leaning.

Comparison with Existing Datasets. · 4 Methods

The paper contrasts its longitudinal, user-centric datasets with prior post-level resources and introduces TSN4PI, whose PIDN combines offline LLM-based style adaptation with online BERT-based ideology detection. The datasets and framework support analysis of political relevance, ideological alignment, and ideological change across X and Truth Social.

  • Comparison with Existing Datasets.: Prior resources are largely post-centric, while some Twitter datasets involve very small user bases, including 79 and 46 users.Examples include Weibo-COV, Koo, iDRAMA-Scored, TweetIntent@Crisis, and News Feed Ranking.
  • Comparison with Existing Datasets.: The X dataset contains full tweet histories for 4,545 users over 16 years, retaining threads and engagement for longitudinal analysis of election discourse and ideological evolution.Its user continuity, temporal span, and scale are presented as advantages over post-level snapshots.
  • Comparison with Existing Datasets.: The X dataset’s monthly activity rises from January 2007 to September 2022, accelerating after 2018 and peaking around the 2020 presidential election.Additional activity peaks occurred during the 2012 and 2016 elections and the early-2020 COVID-19 outbreak.
  • 3.3 Truth Social Dataset: Truth Social was added as a platform launched in February 2022 after Donald Trump’s January 2021 suspension, with functionality closely mirroring X.Truth Social uses “Truths” and “ReTruths” in place of “Tweets” and reposts.
  • 3.3 Truth Social Dataset: The Truth Social dataset covers over 454,000 users and more than 823,000 Truths, including complete post histories for the 65,536 most active users.Its primary exploratory resources include the Truths and Users subtables.
  • 4 Methods: TSN4PI comprises two modules, PIDN and PIPN, addressing political ideology detection and future ideological-shift prediction.The supplied methods section explicitly introduces the framework as consisting of both modules.
  • 4.1 Political Ideology Detection Network: PIDN first filters non-political content and then estimates political posts’ positions on the ideological spectrum using UDA and text style transfer.This formulation separates political relevance detection from ideology-score estimation.
  • 4.1 Political Ideology Detection Network: The detection pipeline uses offline LLM-based text style transfer to align formal news with social-media language, followed by an online BERT-based classifier.BERT is selected for inference speed and resource efficiency, while LLMs are reserved for the offline TST stage.

LLMs for Text Style Transfer. · Unsupervised Domain Adaptation for Distribution Alignment.

The framework uses LLM-based style transfer to narrow stylistic differences between source news and target social-media text, then applies unsupervised domain adaptation to align residual distributions without labeled target data. It evaluates MMD and KL-based alignment within a composite objective that balances task losses with cross-domain representation alignment.

  • LLMs for Text Style Transfer.: LLM-based text style transfer narrows the stylistic gap between source-domain news and target-domain social-media texts.Outputs are compared with source news and target Truth Social posts using four BERT-based pretrained language models and UMAP projections.
  • Unsupervised Domain Adaptation for Distribution Alignment.: Residual distributional shifts remain after style transfer, motivating an optional unsupervised domain adaptation module that does not require labeled social-media data.The primary alignment method is Maximum Mean Discrepancy, which measures distribution divergence in a reproducing kernel Hilbert space without paired or balanced samples.
  • Unsupervised Domain Adaptation for Distribution Alignment.: MMD uses a Gaussian RBF kernel to map samples into a space where cross-domain similarity can be measured and optimized.The kernel is defined as k(x,y) = exp(−||x−y||^2/(2σ^2)).
  • Unsupervised Domain Adaptation for Distribution Alignment.: KL divergence provides an alternative alignment method for model outputs interpretable as probability distributions, including softmax scores or normalized embeddings.It explicitly aligns posterior distributions between source prediction P and target prediction Q.
  • Unsupervised Domain Adaptation for Distribution Alignment.: MMD and KL are evaluated for robustness because KL assumes stronger probability-space alignment and is more suitable for well-calibrated output distributions.The two metrics address differing domain conditions and alignment assumptions.
  • Unsupervised Domain Adaptation for Distribution Alignment.: A composite loss jointly optimizes political-relevance classification, ideology regression, and selected domain alignment, with nonnegative weights satisfying α + β ≤ 1.Disabling UDA removes the alignment term; embedding-distribution alignment is intended to improve generalization to informal, unlabeled social-media content while preserving task learning.

Temporal Graph Construction. · Temporal Graph Modeling.

The framework models spatio-temporal interactions with JODIE, APAN, and TGN, then trains temporal node representations through self-supervised link prediction. Classification and regression heads use these temporal embeddings to predict users’ future political ideologies.

  • Temporal Graph Construction.: Temporal Graph Construction. evaluates JODIE, APAN, and TGN to capture spatio-temporal dynamics in interaction data.JODIE uses coupled RNNs to update user and item embeddings, while a projection operator estimates future user embeddings through time-conditioned scaling.
  • Temporal Graph Modeling.: Temporal Graph Modeling. JODIE’s projection incorporates elapsed-time context through w_Δt.The elapsed-time encoding supplies temporal context for future embedding estimation.
  • Temporal Graph Modeling.: Temporal Graph Modeling. APAN asynchronously updates embeddings with an attention-based encoder and a mailbox of recent neighbor interactions.Each interaction uses the last state together with stored recent-neighbor information.
  • Temporal Graph Modeling.: Temporal Graph Modeling. TGN maintains time-evolving memory, generates interaction messages, and updates node states through a memory function.Final node embeddings are computed using temporal attention over neighbors.
  • Temporal Graph Modeling.: Temporal Graph Modeling. self-supervised temporal link prediction trains the TGNN to forecast future edges from historical graph sequences.For candidate node pairs, link scores use time-specific embeddings, a learnable projection matrix W, and the sigmoid function σ(·).
  • Temporal Graph Modeling.: Temporal Graph Modeling. produces temporally contextualized node embeddings h_t^n for downstream ideology-prediction heads.These representations are generated by the self-supervised link-prediction stage.
  • Temporal Graph Modeling.: Temporal Graph Modeling. attaches classification and regression heads to temporal node embeddings to predict users’ future political ideologies.The passage specifies both head types but does not report quantitative results here.

Node Classification and Regression. · 5 Experiments

TSN4PI jointly models users’ political relevance and ideological intensity from historical posts, then evaluates PIDN detection and PIPN prediction across temporal and static graph settings. The experiments assess backbone scaling, style transfer, domain adaptation, and temporal graph modeling.

  • Node Classification and Regression.: Political relevance is derived from the majority relevance label among a user’s historical posts up to time t.PIDN-derived post labels are aggregated into a node-level relevance target.
  • Node Classification and Regression.: A user’s ideology target is the mean PIDN ideology score over politically relevant historical posts.Only posts labeled relevant contribute to the regression target.
  • Node Classification and Regression.: A unified loss jointly optimizes classification and regression, with λ∈[0, 1] balancing the two tasks.The multi-task objective learns temporally grounded representations of ideological relevance and intensity.
  • 5 Experiments.: The experiments validate ideology detection and prediction through setup, PIDN evaluation, and subsequent PIPN analysis.Datasets, baselines, and implementation details are introduced before module-specific evaluations.
  • 5.1.1 Datasets and Data Preparation.: PIDN training combines style-transferred political content with an equal number of politically irrelevant daily-life tweets.Irrelevant tweets receive relevance 0 and placeholder ideology scores excluded from regression.
  • 5.1.1 Datasets and Data Preparation.: PIPN experiments use Truth Social and X temporal graphs built from PIDN-identified and scored political posts, while static baselines discard temporal information.Static GNN graphs aggregate temporal interactions into weighted graphs.
  • 5.1.2 Models, Baselines, and Evaluation Metrics.: PIDN evaluation covers Qwen2.5 zero-shot scaling from 0.5B to 72B parameters, Qwen2.5-7B few-shot settings, and TST/UDA ablations.Style transfer is assessed with embedding cosine similarity and LLM-as-a-Judge scores normalized to [0, 1].
  • 5.1.2 Models, Baselines, and Evaluation Metrics.: PIPN compares TGN, JODIE, and APAN against static GCN and GAT baselines for temporal interaction forecasting and downstream ideology prediction.Downstream prediction combines political node classification with ideological-score regression.

1) self-supervised link prediction,

The evaluation shows that TSN4PI combines highly accurate, efficient ideology detection with temporal graph modeling that improves link prediction and ideology analysis across Truth Social and X. Component studies identify twhin-bert-base and TST as important design choices, while static graph baselines consistently lag temporal models.

  • PIDN performance: PIDN achieved 99.25% classification accuracy, 0.0048 regression MSE, and 0.069 RMSE, demonstrating precise ideology detection and scoring.The benchmark used five repeated few-shot evaluation settings with randomly drawn, class-balanced held-out examples.
  • LLM comparison: Qwen2.5-72B-Instruct achieved the highest zero-shot Macro F1-score of 0.953, while 7B–32B models reached precision up to 0.992 but recall down to 0.827.The scaling trend was positive but non-monotonic, with mid-sized models showing conservative prediction behavior.
  • PIDN efficiency: PIDN reached 4.39 ms inference latency, supporting efficient deployment relative to larger Qwen2.5 language-model baselines.The comparison used a twhin-bert-base backbone with approximately 0.3B parameters.
  • Architecture and ablation: twhin-bert-base produced the best backbone results, with 97.79% accuracy for political relevance classification and 0.0606 MSE for ideology regression.The backbone comparison excluded TST and UDA and used fixed hyperparameters.
  • Architecture and ablation: Adding TST produced the most dramatic ablation improvement, while the selected hyperparameters α=0.4 and β=0.3 achieved 97.79% accuracy and 0.067 MSE.The ablation incrementally added TST, KL-divergence UDA, and MMD UDA to the twhin-bert-base backbone.
  • Temporal graph comparison: On Truth Social, JODIE led link prediction with 99.73% AP and 99.69% AUC, whereas TGN led ideology classification with 72.42% accuracy and tied APAN at 0.036 regression MSE.Static baselines substantially underperformed the temporal graph neural networks.

6 Results and Findings

The results validate TSN4PI as an integrated two-stage framework linking textual ideological annotation with temporal network dynamics. PIDN-based analysis reveals differing ideological landscapes across Twitter and Truth Social, supporting investigation of ideological shifts, echo chambers, and polarization.

  • Framework validation: TSN4PI integrates PIDN for reliable ideological annotation with TGNN-based PIPN for modeling temporal interaction and ideological-structure dynamics.The two stages connect micro-level textual expression with macro-level network dynamics.
  • Cross-platform ideological landscapes: Visualizations examine ideological shifts across platforms and provide insights into echo chambers and political polarization.The analysis covers social media post content across different levels and objectives.
  • Cross-platform ideological landscapes: PIDN maps distinct platform landscapes: Twitter shows a modest leftward tilt, whereas Truth Social skews markedly rightward.The distributions align with prevailing public perceptions despite differences in dataset size.

Different Social Media Platforms exhibit distinct overarching Political Ideologies.

On X (Twitter), highly volatile users followed complex, multidirectional ideological trajectories rather than consistent polarization, including moderation toward or across the political center. Users with initially centrist positions also showed substantial volatility, indicating that centrism did not ensure future stability.

  • User-level trajectory analysis: The analysis identified the 20 X users with the widest spread between their highest and lowest yearly mean ideology scores, requiring at least 10 tweets annually and 50 overall.These criteria were used to support robust estimation of highly volatile users’ ideological trajectories.
  • Non-polarizing trajectories: Many initially right-leaning users moved toward the ideological center, with several crossing into left-leaning territory over time.The observed pattern contradicted the classical echo-chamber hypothesis and suggested ideological moderation or de-radicalization rather than partisan amplification.
  • Non-polarizing trajectories: Users in the high-variability cohort fluctuated widely across the political spectrum, and an initial centrist position did not predict future stability.A complementary case study examined whether users with highly variable predicted ideology scores exhibited identifiable individual-level dynamics.

Qualitative analysis reveals diverse, non-polarizing ideological trajectories. · Comparative Thematic Analysis Reveals Divergent Discursive Arenas. · 7 Conclusion

The paper finds that users follow diverse ideological trajectories rather than uniformly radicalizing, while platforms differ between interconnected competing frames and a high-conflict, right-shifted arena. TSN4PI unifies language processing and temporal graph modeling to track these dynamics and contributes two large-scale datasets.

  • Qualitative analysis reveals diverse, non-polarizing ideological trajectories.: The three most volatile user cases do not support the classical echo chamber hypothesis of unidirectional radicalization.Users instead exhibit nuanced ideological shifts, including moderation, continuity, and changing rhetorical intensity.
  • Qualitative analysis reveals diverse, non-polarizing ideological trajectories.: Another user changed topics while maintaining a stable ideological stance, reflecting continuity rather than radicalization.The user expressed skepticism toward institutions through constitutional-violation claims and rejection of public-health guidance.
  • Qualitative analysis reveals diverse, non-polarizing ideological trajectories.: Together, the cases challenge the universality of the polarization narrative.The analysis therefore presents ideological evolution as heterogeneous rather than uniformly polarizing.
  • Comparative Thematic Analysis Reveals Divergent Discursive Arenas.: Twitter exhibits a shared agenda with competing frames, as political groups discuss common high-salience topics but interpret them differently.Examples include divergent framing of impeachment and January 6th, with distinct topic terms across ideological groups.
  • Comparative Thematic Analysis Reveals Divergent Discursive Arenas.: Truth Social forms a high-conflict ideological arena with a right-shifted center, where partisan narratives become mainstream and rhetorical intensity rises.Themes labeled far-right on Twitter appear under Truth Social’s Centrist category, alongside platform-specific claims and intensified framing.
  • Comparative Thematic Analysis Reveals Divergent Discursive Arenas.: Across platforms, Twitter is fragmented but interconnected, whereas Truth Social operates as a self-reinforcing enclave that amplifies division.The contrast concerns how competing frames coexist on Twitter versus how moralized, reactive rhetoric reinforces polarization on Truth Social.
  • 7 Conclusion: TSN4PI integrates natural language processing with temporal graph neural networks to track political ideologies on social media.The framework bridges news and social-content domains, models dynamic user interactions, and contributes two large-scale datasets from X and allsides.com.

A Appendix … A.2 Implementation Details

The appendix describes privacy-preserving data handling, preprocessing and normalization for the datasets, ideological language patterns in news titles, and implementation settings for model training and inference.

  • A.1 Details of the Datasets: Public social-media posts were anonymized by replacing personal identifiers with numeric IDs, while private user data were excluded and released datasets were dehydrated.These measures addressed privacy and source terms for X and Truth Social data.
  • A.1.2 The allsides.com Dataset.: AllSides preprocessing retained English articles and normalized original ideology scores from [−6, 6] to the [0, 1] interval.The normalized scale supports uniform interpretation during model training and evaluation.
  • A.1.2 The allsides.com Dataset.: Left-leaning news emphasized the Capitol riot, Supreme Court and COVID, whereas right-leaning news emphasized Elon Musk, Paul Pelosi, criminal committees and vaccine mandates.Center media reported on topics associated with both sides, reflecting differing issue focus and language use.
  • A.1.2 The allsides.com Dataset.: Tweets were filtered to English using the fasttext-langdetect library to maintain linguistic consistency with the presidential-election analysis.The filtering targeted the United States presidential-election context.
  • A.1.3 The X (Twitter) Dataset.: Missing or misaligned Truth Social timestamps were supplemented and corrected through web scraping for accessible posts.The correction applied to some “Truths” in the raw dataset.
  • A.2 Implementation Details: Models used a 90/10 train-test split; PIDN training lasted 3 epochs with a 2×10−5 learning rate and 0.01 weight decay.The TST configuration used Llama-3.1-8B-Instruct with temperature 0.7, top_p 0.95 and a 512-token maximum.
  • A.2 Implementation Details: The implementation used vllm for text generation, TGL for temporal graph neural networks, PyTorch Geometric for static graphs and PyTorch 2.4.0 for training and inference.

A.3 Prompt Details · Prompt for LLM-based method.

The prompt template uses tool calling to produce structured political-relevance and ideological-leaning analyses of tweets. It specifies relevance classification, conditional ideological scoring, step-by-step reasoning, and a JSON tool-call output format, alongside a separate text-style-transfer prompt for converting news into concise tweets.

  • A.3 Prompt Details: Tool calling enables the model to reason about a tweet’s political relevance and leaning before outputting structured results.The template is designed to generate relevance and scores for posts.
  • Prompt for LLM-based method.: The system prompt assigns an expert U.S.-politics analyst role and requires analysis through the record_political_analysis tool.The tool records the political analysis of a tweet after reasoning.
  • Prompt for LLM-based method.: Relevance is encoded as ’1’ for politically relevant U.S. tweets and ’0’ otherwise, while score is provided only for relevant tweets.The score ranges from 0.0 for left-leaning to 1.0 for right-leaning tweets.
  • Prompt for LLM-based method.: The prompt requires clear step-by-step reasoning, escaped double quotes, and direct use of the analysis function for output.Reasoning is a required string parameter, whereas score is optional and conditional on relevance=='1'.
  • Prompt for LLM-based method.: The example labels an infrastructure-bill tweet as politically relevant and assigns it a centrist score of 0.5.The example cites the Senate, administration, and balanced criticism as evidence for the classification.
  • Prompt for LLM-based method.: The input uses {tweet_text}, and the model should directly return tool-call arguments as a single well-formed JSON string.A separate text-style-transfer prompt converts news articles into concise tweets for training.

Prompt for Text Style Transfer. · Prompt for Text Style Transfer Evaluation.

The text-style-transfer prompt rewrites news reports as concise English tweets with social-media conventions and a required output token. A separate LLM judge scores tweet naturalness and key-information preservation using structured tool calls.

  • Prompt for Text Style Transfer.: The news-to-tweet system prompt directs rewriting a news report to capture its core message as a tweet.
  • Prompt for Text Style Transfer.: Generated tweets must be concise, stay within 280 characters, use abbreviations and hashtags, mention relevant accounts, and end with </Tweet>.
  • Prompt for Text Style Transfer.: The evaluation prompt uses an LLM-as-a-judge to assess generated-tweet relevance, quality, style, and semantic preservation relative to the original article.
  • Prompt for Text Style Transfer Evaluation.: The judge assigns two integer scores from 0–10: style_score for tweet naturalness and semantic_score for key-information preservation.
  • Prompt for Text Style Transfer Evaluation.: The rubric maps performance from 10 for flawless style and complete facts to 0 for gibberish or unrelated content.
  • Prompt for Text Style Transfer Evaluation.: The judge must reason about style and semantics, then call grade_tweet with required integer style_score and semantic_score parameters.

A.4 UDA Complexity Analysis … Summary and Implications.

The UDA complexity analysis compares MMD and KL divergence in theoretical runtime, FLOPs, and GPU implementation characteristics. MMD has quadratic dependence on batch size, whereas KL divergence scales linearly and is more suitable for large-batch training.

  • A.4 UDA Complexity Analysis: The analysis evaluates MMD and KL divergence across theoretical time complexity, floating-point operations, and practical GPU implementation characteristics.It assumes source and target embedding batches containing N samples of dimension D.
  • Maximum Mean Discrepancy (MMD).: MMD constructs three N×N kernel matrices, K_XX, K_YY, and K_XY.These matrices capture pairwise relationships between the source and target embedding sets.
  • Maximum Mean Discrepancy (MMD).: O(N^2D) is the time complexity for forming one Gaussian RBF kernel matrix in MMD.Computing each pairwise squared Euclidean distance costs O(D), and all N×N pairs must be evaluated.
  • Kullback-Leibler (KL) Divergence.: O(ND) is the time complexity of KL divergence over distribution-like outputs P and Q with shape (N, D).The computation applies element-wise division, logarithm, and multiplication before summation.
  • Summary and Implications.: The choice between MMD and KL divergence introduces a critical trade-off in computational efficiency.Table 5 summarizes the computational characteristics of the two domain-adaptation methods.
  • Summary and Implications.: MMD’s quadratic complexity restricts its application to smaller batch sizes, while KL divergence remains efficient for large-batch training.Batch sizes must therefore be managed carefully with MMD, whereas KL divergence provides a more scalable computational alternative.
Loading 2608.17987v1…