PaperlayerView exact revision on arXiv

Source-linked AI summary

A Multi-Branch Feature Fusion Approach for Health Misinformation Detection and Propagation

Mkululi Sikosana, Sean Maudsley-Barton, Oluwaseun Ajao

arXiv:2609.00403v1cs.LGcs.SI

TL;DR

Health misinformation detection needs scalable methods that also address diffusion when engagement supervision is incomplete. This paper combines transformer semantics with rhetorical, stance, ELM, and TPB proxies in a multi-branch framework and adds CPS as an engagement-independent diffusion-propensity index. Across three datasets, fusion generally improves detection and supports interpretable diffusion-oriented analysis, while CPS still requires human-centred validation.

  • Problem

    Manual fact-checking does not scale to fast-moving OSN misinformation, and engagement traces are unavailable uniformly across datasets, limiting propagation modelling.

  • Method

    A multi-branch multi-task model fuses transformer, rhetorical, stance, and ELM- and TPB-aligned text features, with CPS computed post hoc from engagement-independent cues.

  • Results

    Across three datasets, multi-perspective fusion strengthens within-dataset detection and supports interpretable diffusion-propensity analysis; semantic features drive performance while psychological features provide complementary gains.

  • Takeaways & Limitations

    The framework links cognitive theory with neural misinformation modelling and supports feature-level interpretation across datasets with differing linguistic and diffusion-risk characteristics.

  • Takeaways & Limitations

    CPS requires validation through human-centred protocols and expert annotation, and the study calls for multilingual, multimodal, statistical, and sensitivity extensions.

Abstract

from arXiv · show

This paper presents a multi-branch fusion framework for detecting and characterising the propagation of health misinformation in online social networks (OSNs). Grounded in the Elaboration Likelihood Model (ELM) and the Theory of Planned Behaviour (TPB), the model fuses transformer-based semantics with rhetorical cues, stance representations, and psychologically motivated proxies in a unified multi-task architecture. In addition to binary classification, we introduce the Cognitive Propagation Score (CPS), an interpretable post-hoc auxiliary score computed from psychologically motivated, text-derived cues capturing argument complexity, emotional intensity, and content-derived virality potential, to support diffusion-risk reasoning when engagement ground truth is incomplete or unavailable. Experiments on three benchmark datasets, Constraint, COVID--19\_FNIR, and Monkeypox, show strong classification performance, achieving ROC--AUC up to 0.9999 on COVID--19\_FNIR, while propagation-oriented ranking achieves near-perfect agreement when engagement-derived supervision is available (Monkeypox, Spearman's $ρ= 0.9952$) and similarly high ranking alignment under proxy-based supervision on COVID--19\_FNIR ($ρ= 0.9954$). Compared with representative literature baselines, the fusion model improves detection on Constraint and COVID--19\_FNIR, while Monkeypox remains more challenging, reflecting domain- and signal-specific differences. Ablation analysis further indicates that psychological and rhetorical branches provide complementary gains beyond semantic embeddings. Overall, the framework bridges cognitive theory and neural modelling to improve transparency and to support scalable misinformation monitoring, with future work required to validate CPS against human-centred diffusion judgements.

I. INTRODUCTION

Health misinformation spreads rapidly through OSNs, creating public-health and institutional risks that manual fact-checking cannot address at scale. The paper responds with a cognitively grounded multi-branch framework combining semantic, rhetorical, stance, and theory-informed text features for detection and diffusion-oriented analysis.

  • Motivation: Manual fact-checking cannot scale to the volume and speed of modern health misinformation outbreaks.During COVID-19, misleading claims circulated rapidly and were associated with unsafe behaviours and reduced confidence in health guidance.
  • Motivation: Transformer models capture contextual language patterns but provide limited insight into why misinformation persuades and diffuses.ELM and TPB motivate more transparent feature design around cognition, affect, attitudes, norms, and perceived control.
  • Research gap: Missing engagement traces across datasets motivate text-derived diffusion signals for ranking and prioritisation when native engagement supervision is unavailable.Some corpora provide engagement metadata while others provide only veracity labels.
  • Approach: The proposed framework fuses transformer embeddings with stance, rhetorical structure, and ELM- and TPB-aligned proxies in a multi-task architecture.The theory-informed proxies are extracted from post text to support feature-level interpretation.
  • Approach: CPS is an engagement-independent post-hoc scalar indicating relative diffusion propensity rather than retweet count.Observed engagement may supervise propagation regression when available, but it does not enter the CPS calculation.
  • Evaluation: The study evaluates the framework across Constraint, COVID-19-FNIR, and Monkeypox benchmarks against classical, transformer, and lightweight neural baselines.Reported references include XLNet+LDA and linguistic-feature SVM on Constraint, DistilBERT and CNN–LSTM results on COVID-19-FNIR, and BERT-based PoxVerifi on mpox claims.

III. MATERIALS & METHODS

The study applies a common preprocessing and feature-engineering pipeline to three misinformation datasets, then supplies scaled multi-branch representations to classification and propagation regression. Raw quantities are retained separately for post-hoc CPS computation, while TTR remains length-dependent despite training-only scaling.

  • Data and pipeline: Three datasets share tokenisation, lowercasing, stop-word removal, punctuation standardisation, and common feature extraction and scaling.The datasets differ in size and annotation richness, but the pipeline preserves methodological consistency.
  • Feature branches: The fusion inputs comprise DistilBERT semantic embeddings, rhetorical features, stance features, and ELM and TPB psychological features.The stated dimensions are 768 for semantic embeddings, 4 each for rhetorical and stance features, 10 for ELM, and 9 for TPB.
  • Psychological features: ELM features are text-derived linguistic proxies rather than direct measurements of elaboration, beliefs, or psychological processing states.Central and peripheral route cues are computed using standard NLP libraries.
  • Central-route features: The central-route representation contains five interpretable features including readability, lexical diversity, sentiment polarity, token count, and average sentence length.The raw TTR is min–max scaled using training-partition parameters before fusion.
  • Feature limitation: TTR scaling aligns numerical ranges but does not remove dependence on text length, so TTR is only an approximate within-corpus lexical-diversity indicator.Length-independent alternatives such as MATTR or MSTTR were not applied.

Peripheral Route Features

The peripheral-route branch represents emotional, stylistic, and heuristic appeal through interpretable text-derived cues. These features complement the broader ELM and TPB proxy design used in the fusion architecture.

  • Peripheral-route representation: The ELM peripheral-route representation comprises five interpretable text-derived features capturing emotional, stylistic, and heuristic appeal.The branch is represented as a five-component vector.
  • Peripheral features: Peripheral cues include exclamation-mark ratio, question-mark ratio, uppercase-letter ratio, all-uppercase word count, and urgency-term ratio.These quantify emphatic punctuation, questioning, capitalisation intensity, shouting language, and predefined urgency terms.
  • Theory-linked proxies: The wider feature design operationalises TPB through textual proxies for attitude, subjective norms, and perceived behavioural control.Examples include affective polarity, group pronouns, normative phrasing, certainty, and directive language.

Perceived Behavioural Control:

The TPB-aligned branch encodes perceived behavioural control and related constructs with linguistic proxies for certainty, evaluation, social referencing, and directives. These features approximate psychological cues rather than directly measuring users’ attitudes or intentions.

  • Measurement boundary: The TPB components are theory-aligned text-derived proxies because the datasets contain no direct self-report measures of attitudes, norms, control, or intentions.They should not be interpreted as definitive evidence of internal psychological states.
  • Attitude-aligned proxies: Attitude-aligned features use sentiment polarity, lexical valence, and modality to represent affective evaluation and expressed commitment.They concern the post topic or claim and do not directly measure attitudes toward misinformation or sharing behaviour.
  • Sentiment caveat: VADER provides an approximate indicator of expressed textual valence, but corpus-specific accuracy could not be calculated without independently validated sentiment annotations.Sarcasm, quotations, implicit evaluations, health terminology, negation, and mixed sentiment can produce misleading scores.
  • Subjective-norm proxies: Subjective-norm proxies capture group pronouns, retweet or mention tokens, and social-comparison expressions.Absent markers are set to zero, preserving a consistent nine-dimensional TPB vector across datasets.
  • Perceived behavioural control: Perceived behavioural control is represented through certainty cues, instructional language, and mobilisation-oriented hashtag directives.Examples include “definitely,” “click here,” “share this,” #WakeUp, and #JoinUs.
  • Scaling and CPS inputs: Engineered ELM and TPB features are scaled to [0, 1], while raw CPS quantities are retained separately for post-hoc computation.The retained quantities include readability, adjective count, VADER sentiment, punctuation count, and character length.

E. THEORETICAL JUSTIFICATION FOR CPS AND TPB

The framework translates ELM and TPB constructs into scalable text-derived proxies within a multi-branch fusion architecture. CPS remains a separate post-hoc index computed from selected raw text-derived quantities rather than the complete psychological feature vectors.

  • E. THEORETICAL JUSTIFICATION FOR CPS AND TPB: ELM- and TPB-related constructs are reformulated as psycholinguistic proxies for the psychological branch of the fusion architecture.The approach addresses the limited scalability of traditional self-report instruments for large social-media corpora.
  • E. THEORETICAL JUSTIFICATION FOR CPS AND TPB: The fused representation aggregates semantic embeddings, rhetorical cues, stance representations, ELM features, and TPB features.The semantic branch uses a 768-dimensional DistilBERT contextual representation.
  • E. THEORETICAL JUSTIFICATION FOR CPS AND TPB: Table 1 defines the engineered features, their operational meanings, dimensionality, and dataset availability for the fused input representation.This inventory provides the reference definition for the feature branches used throughout the paper.
  • E. THEORETICAL JUSTIFICATION FOR CPS AND TPB: Stance is represented using a 4-way encoding over support, denial, query, and comment.This encoding forms part of the stance branch within the fused representation.
  • E. THEORETICAL JUSTIFICATION FOR CPS AND TPB: CPS is computed separately from selected raw text-derived quantities and is not generated from the shared hidden representation.The learned fusion outputs and post-hoc CPS therefore have distinct computational roles.

LEARNED OUTPUT HEADS AND POST-HOC CPS

The model learns separate classification and propagation-regression heads, while CPS is a non-trainable, engagement-independent post-hoc index. CPS combines weighted text-derived proxies for complexity, emotional intensity, and content-derived virality potential.

  • LEARNED OUTPUT HEADS AND POST-HOC CPS: The two learned objectives are binary misinformation classification and propagation estimation through separate output heads.The shared representation feeds both heads, whereas CPS is computed afterward and does not participate in optimisation.
  • LEARNED OUTPUT HEADS AND POST-HOC CPS: The classification head outputs a misinformation probability through a sigmoid-activated linear layer and binary cross-entropy objective.This head predicts whether an input tweet contains misinformation.
  • LEARNED OUTPUT HEADS AND POST-HOC CPS: The regression head estimates a continuous propagation target using observed engagement for Monkeypox and a transferred propensity proxy for Constraint and COVID–19_FNIR.Propagation estimation is methodologically distinct from CPS.
  • LEARNED OUTPUT HEADS AND POST-HOC CPS: CPS combines argument complexity, emotional intensity, and content-derived virality potential using fixed post-hoc weights of 0.2, 0.4, and 0.4.The components use Flesch–Kincaid grade level, adjective-token count, sentiment, punctuation count, and character length.
  • LEARNED OUTPUT HEADS AND POST-HOC CPS: Raw CPS values are not constrained to [0, 1], so their magnitude is dataset-dependent and may include occasional small negative values.The raw scale is retained to preserve reported tail behaviour.
  • LEARNED OUTPUT HEADS AND POST-HOC CPS: CPS excludes observed engagement variables and is intended as an auxiliary ranking signal when engagement information is incomplete, delayed, or absent.It is not a third neural output head and does not estimate retweet, like, reply, or view counts.

G. PROPENSITY-TO-SPREAD PROXY CONSTRUCTION

The propensity-to-spread proxy transfers an engagement-supervised logistic model from Monkeypox to datasets lacking native engagement metadata. It supports relative ranking but is not an absolute estimate of realised platform virality.

  • G. PROPENSITY-TO-SPREAD PROXY CONSTRUCTION: Monkeypox supplies the source-domain engagement supervision because Constraint and COVID–19_FNIR lack equivalent engagement fields.The source target uses available retweet, like, and reply counts.
  • G. PROPENSITY-TO-SPREAD PROXY CONSTRUCTION: High engagement is defined as the upper 10% of the Monkeypox engagement distribution using a training-partition percentile threshold.The threshold is applied unchanged to held-out data to avoid target leakage.
  • G. PROPENSITY-TO-SPREAD PROXY CONSTRUCTION: The logistic propensity model uses standardized text-derived features that can be computed consistently across all three datasets.The feature set includes length, sentiment, affect, punctuation, capitalization, URL or evidence, hedging or modality, stance, and rhetorical cues.
  • G. PROPENSITY-TO-SPREAD PROXY CONSTRUCTION: The fitted model maps a linear decision score through a sigmoid, with higher propensity values indicating greater similarity to high-engagement Monkeypox posts.The coefficients and intercept are learned from Monkeypox.
  • G. PROPENSITY-TO-SPREAD PROXY CONSTRUCTION: The frozen feature definitions, scaler, coefficients, and intercept are transferred to Constraint and COVID–19_FNIR without target-dataset engagement labels.The resulting scores are transferred propensity-to-spread estimates rather than observed engagement.
  • G. PROPENSITY-TO-SPREAD PROXY CONSTRUCTION: The transferred score supports ordinal ranking but is not assumed to recover missing engagement counts or provide calibrated absolute virality probabilities.Its transfer assumes that some Monkeypox-associated text cues remain informative across related health-misinformation datasets.
  • G. PROPENSITY-TO-SPREAD PROXY CONSTRUCTION: The propensity proxy and CPS are complementary diffusion-oriented signals with different construction procedures.The proxy is learned under Monkeypox engagement supervision, whereas CPS is engagement-independent and post hoc.

H. LOSS FUNCTIONS

The model jointly optimizes misinformation classification and propagation regression using a weighted total loss. Propagation supervision depends on dataset availability, while CPS remains outside the training objective.

  • H. LOSS FUNCTIONS: The training architecture has two learned objectives: binary misinformation classification and propagation regression.CPS is calculated post hoc and does not contribute a training loss.
  • H. LOSS FUNCTIONS: The classification objective uses binary cross-entropy, while propagation estimation uses mean squared error.The two losses correspond to the classification and regression heads.
  • H. LOSS FUNCTIONS: Propagation regression uses observed engagement intensity for Monkeypox and the transferred propensity-to-spread proxy for Constraint and COVID–19_FNIR.The regression output is interpreted relative to each dataset’s supervision target rather than universal platform engagement.
  • H. LOSS FUNCTIONS: Fixed validation-selected weights λ_cls and λ_reg balance the classification and regression objectives in the total loss.The loss weights are positive hyperparameters.
  • H. LOSS FUNCTIONS: Training uses AdamW, a linear learning-rate scheduler with warm-up, staged DistilBERT unfreezing, and early stopping.Tweets are truncated to 128 tokens, engineered features are min–max scaled, and data are split into stratified 60/20/20 train–validation–test subsets.

J. INFERENCE WORKFLOW

The inference workflow combines a unified evaluation protocol with learned classification and propagation outputs, while CPS is computed separately from retained raw text-derived cues. Evaluation distinguishes implemented experiments from literature context and adapts propagation targets to engagement availability.

  • J. INFERENCE WORKFLOW: The model produces learned classification and regression outputs from a shared representation, while CPS is calculated separately from retained raw psychology-aligned cues.CPS requires neither the hidden-state representation nor observed engagement variables.
  • J. INFERENCE WORKFLOW: The implemented experiments use a unified protocol, including a stratified 60/20/20 split, a fixed random seed of 42, and model selection by the lowest validation Ltotal.These choices are intended to reduce variance from data partitioning and optimisation.
  • J. INFERENCE WORKFLOW: Literature-reported baselines serve as contextual references because published studies differ in splits, preprocessing, feature access, and tuning regimes.The paper avoids treating such results as strictly protocol-matched comparisons.
  • J. INFERENCE WORKFLOW: Propagation supervision uses engagement-derived targets when metadata exist and propensity-score proxies when engagement fields are unavailable.This preserves a dataset-specific propagation objective while avoiding direct reliance on unavailable engagement ground truth.
  • J. INFERENCE WORKFLOW: CPS is characterised descriptively rather than evaluated as a prediction task, using distributional statistics such as mean, standard deviation, extrema, and log10 maximum.Propagation prediction uses error, ranking, and distributional metrics, whereas CPS is deterministic and post hoc.

A. CLASSIFICATION PERFORMANCE

Classification is strongest on COVID-19_FNIR and Constraint and weaker on Monkeypox, while propagation performance varies with target construction and threshold calibration. Ablations show semantic features drive performance, with psychological, rhetorical, and stance cues adding dataset-dependent complementary information; CPS distributions are highly dataset-specific.

  • A. CLASSIFICATION PERFORMANCE: Approximately 80% accuracy on Monkeypox is lower than the strong classification performance reported for COVID-19_FNIR and Constraint.The paper attributes the challenge plausibly to partial truths and emotive framing that blur class boundaries, while cautioning that COVID-19_FNIR ceiling metrics are within-dataset results.
  • B. ABLATION STUDY: CONTRIBUTION OF FUSION: Removing the semantic branch produces the largest F1-score and ROC-AUC declines across datasets, identifying DistilBERT representations as the primary performance driver.The ablation removes one branch at a time under identical hyperparameters and splits.
  • B. ABLATION STUDY: CONTRIBUTION OF FUSION: The psychological branch consistently improves discrimination, especially on Monkeypox and Constraint, whereas rhetorical and stance branches matter more on Constraint than on the other datasets.These patterns indicate that behavioural, affective, style, and stance cues contribute differently across corpora.
  • C. PROPAGATION PREDICTION: COVID-19_FNIR achieves the lowest propagation errors, a Top-5 hit rate of 1.000, and low KL divergence under its chosen proxy-defined virality setup.Constraint has MSE approximately 0.0019 but viral recall 0, while Monkeypox has strong rank ordering but viral recall of 62.6%.
  • C. PROPAGATION PREDICTION: Constraint propagation evaluation uses a propensity proxy rather than native engagement and shows threshold miscalibration for extreme outliers, including viral recall of 0.Its higher KL divergence of 2.53 indicates pronounced distributional mismatch.
  • D. COGNITIVE PROPAGATION SCORE (CPS): Raw CPS distributions differ markedly: Constraint has a heavy upper tail, whereas Monkeypox has a tighter range with maximum approximately 57 and near-zero scores.Constraint’s maximum is approximately 1,258, and high-CPS items are treated as ranking candidates rather than posts with known realised reach.

E. DATASET-LEVEL ANALYSIS: CLASSIFICATION,

Across the three datasets, classification is strongest on COVID–19_FNIR and Constraint, while propagation alignment depends on engagement availability and dataset-specific diffusion structure. CPS profiles and ablations indicate complementary psychological, rhetorical, and semantic signals, but Monkeypox and outlier-heavy regimes remain more challenging.

  • Constraint: Constraint reaches 97.81% accuracy and 97.74% F1, but propagation correlation is 0.2272 with KL divergence of 2.5319.This divergence indicates that strong classification does not ensure accurate virality retrieval under a global threshold in an outlier-heavy regime.
  • CPS profiles: CPS is heavy-tailed on Constraint, with mean 23.07 and maximum 1257.78, but bounded on Monkeypox at maximum 57.02.These profiles correspond to rare extreme cascades in Constraint and fewer extreme outliers in Monkeypox.
  • Monkeypox: Monkeypox records F1 = 80.82% alongside Spearman’s ρ ≈0.9952, high viral precision, moderate recall, and higher KL divergence.The model orders cases well while retaining distributional mismatch under engagement-derived supervision.
  • COVID–19_FNIR: COVID–19_FNIR achieves F1 = 99.32% and ROC-AUC = 0.9999, with proxy-target ranking ρ = 0.9954 and MSE = 0.0003.Its propagation results align with a propensity-to-spread proxy rather than observed platform engagement.
  • Fusion effects: Ablations show that removing any branch reduces performance, with the largest drops after removing semantics and additional consistent reductions after removing psychological features.Stance and rhetorical branches provide smaller but systematic gains beyond contextual semantics.
  • Interpretation and limits: CPS is intended as an auxiliary diffusion-risk signal rather than a replacement for engagement-based ground truth, and it still lacks human-centred validation.Future evaluation is proposed through share-intent ratings, expert review, or controlled attention studies.
  • Baseline comparison: The model exceeds reported baselines on Constraint and COVID–19_FNIR but underperforms on Monkeypox by 9.75 F1 points relative to Mohbey et al.Literature comparisons are contextual because published protocols may differ.

V. DISCUSSION

The discussion interprets fusion as combining complementary evidence streams for detection and diffusion-risk analysis, while stressing that propagation targets, dataset structure, and evaluation protocols shape the conclusions. CPS improves interpretability as an engagement-independent auxiliary cue, but cross-dataset deployment requires calibrated, dataset-aware use.

  • Discussion: The model integrates psychological, linguistic, stance, rhetorical, and contextual cues to improve classification and interpret model behaviour across datasets.Its intended scope includes diffusion-oriented analysis under the supervision available for each dataset.
  • Complementary branches: Removing the semantic branch causes the largest ablation drop, while psychological, stance, and rhetorical branches add complementary information beyond contextual semantics.Psychological features produce consistent non-trivial F1 reductions when removed.
  • Propagation supervision: Propagation results must distinguish observed engagement targets from propensity-to-spread proxies, because proxy-supervised outputs estimate relative diffusion risk rather than platform engagement.This boundary applies especially to Constraint and COVID–19_FNIR.
  • Dataset regimes: Monkeypox shows strong rank ordering but higher KL divergence, whereas Constraint combines strong detection with weak propagation alignment in a tail-dominated proxy regime.The evaluation design and target distribution therefore affect whether ranking or distributional matching is more informative.
  • CPS interpretation: CPS operationalises argument complexity, emotional intensity, sentiment, punctuation, and text length as interpretable psychology-aligned cues.It is calculated post hoc and independently of observed engagement, regression targets, and shared hidden representations.
  • CPS distributions: CPS distributions vary by dataset, with a skewed Constraint upper tail, a more bounded COVID–19_FNIR range, and a tighter Monkeypox band including near-zero scores.These patterns support dataset-specific interpretations of diffusion propensity rather than a universal score scale.
  • Generalisability: The model improves over reported baselines on Constraint and COVID–19_FNIR but underperforms on Monkeypox, motivating dataset-aware architectural adaptation.Short, syntactically sparse posts and bounded diffusion regimes may reduce the current fusion branches’ advantage.
  • Operational implications: In proxy-supervised settings, ranking and prioritisation are more defensible than absolute thresholding, while engagement-supervised settings may require local calibration.Strong rank agreement can coexist with distributional mismatch.

F. LIMITATIONS & FUTURE DIRECTIONS

The study identifies scope and validation boundaries for its diffusion-oriented findings, including untested multilingual transfer, human-centred CPS validation, target-definition sensitivity, proxy dependence, and challenges with indirect language. Future work therefore targets human validation, multilingual extension, statistical sensitivity analysis, multimodal content, and stronger deployment safeguards.

  • Scope and validation limitations: Multilingual and cross-cultural generalisability was not evaluated because the rhetorical and psychological proxies use English-centric resources and norms.Future work proposes multilingual pretraining, domain adaptation, and cross-cultural robustness checks.
  • Scope and validation limitations: CPS operationalises a narrower set of psychology-aligned textual cues and has not yet been validated through human-centred evaluation.Proposed validation includes share-intent ratings, public-health reviewer assessments, and controlled attention or deliberation studies.
  • Evaluation sensitivity: Diffusion-oriented evaluation is sensitive to target definitions, virality thresholds, top-k settings, and partitioning protocols.The paper recommends repeated runs, systematic sensitivity analyses, and resampling-based uncertainty estimates.
  • Evaluation sensitivity: Constraint and COVID–19_FNIR use deterministic text-derived propensity-to-spread targets, so agreement validates proxy-defined risk rather than real-world propagation.External validation against observed diffusion traces is required before interpreting these scores as engagement-equivalent estimates.
  • Scope and validation limitations: Subtle pragmatic phenomena, including sarcasm, euphemism, implicit persuasion, and culturally coded phrasing, remain challenging for the framework.The paper points to richer context modelling, conversation-level structure, and targeted annotation as possible responses.
  • Future directions: Future work prioritises human-centred CPS validation, multilingual and multimodal extensions, statistical testing, and sensitivity analysis across targets and thresholds.These directions are presented as prerequisites for broader validation and deployment.
Loading 2609.00403v1…