Memory-Guided Multi-View Multi-Domain Fake News Detection

1 INTRODUCTION

• News domains are correlated, but existing fake-news methods often focus on a single domain, such as politics.The authors argue that simultaneously modeling correlated domains is valuable for real-world platforms.Source: §1 · p. 1
• M3FEND addresses domain shift and domain labeling incompleteness, two challenges identified for multi-domain fake news detection.Domain shift reflects differences in words, emotions, and styles; a story assigned to one channel can relate to multiple domains.Source: §1 · pp. 1–2
• The framework models semantics, emotion, and style, then uses a Domain Adapter to aggregate discriminative cross-view representations for different domains.A Multi-head Adaptive Cross-view Interactor learns cross-view interactions, while domain information guides the adapter.Source: §1 · p. 2
• The Domain Memory Bank combines domain-characteristics memory with per-domain event memories to enrich domain information for news pieces.Similarity between a news piece and each domain's event memory supplies a distribution used to guide the Domain Adapter.Source: §1 · p. 2
• M3FEND improves detection performance in most domains and is reported effective in offline experiments and online tests.The paper also reports deployment in an online detection system handling millions of news pieces daily.Source: §1 · p. 2

2.1.1 Word Usage.

• Different news domains show significant differences in their frequently used words.The analysis visualizes domain word frequencies with word clouds.Source: §2.1.1 · p. 2

2.1.2 Writing Style.

• Style features differ between fake and real news, and their usefulness varies by domain.Logic is described as useful for detecting fake news in Science but not Finance.Source: §2.1.2 · p. 2, Figure 2 · p. 3

2.1.3 Publisher Emotion.

• Publisher-emotion features that distinguish fake news differ across domains; Politics fake news shows more disgust, while Health fake news shows more anger.The figure presents emotion-feature heatmaps for fake and real news across nine domains.Source: §2.1.3 · p. 3, Figure 2 · p. 3

2.2 Domain Labeling Incompleteness

• Domain labeling is incomplete because stories assigned to one domain may relate to several, and domain boundaries are difficult to distinguish from content.TextCNN and BiGRU activations overlap across domains, especially for Society, and the domain-classification accuracy is described as unsatisfying.Source: §2.2 · p. 3

3 MODEL

• The framework extracts multi-view representations, enriches domain information with a memory bank, adapts cross-view evidence to domain discrepancies, and predicts fake versus real.Source: §3 · p. 3

3.1 Problem Definition

• Each news item is encoded as token embeddings plus numerical emotion and style features, with a binary fake/real label and one manually assigned domain label.Source: §3.1 · pp. 3–4

3.2 Multi-view Extractor

• Three extractors encode semantics, emotion, and style; TextCNN handles semantic text, while MLPs encode emotional and stylistic features.Source: §3.2 · p. 4
• Multi-channel extractors learn multiple representation subspaces, while a multi-head adaptive interactor learns weighted cross-view combinations without enumerating every combination.Source: §3.2 · pp. 4–5

3.3.1 Domain Characteristics Memory.

• Each Domain Characteristics Memory unit is learned only from training samples in its corresponding domain and represents that domain's characteristics.Source: §3.3.1 · p. 5

3.3.2 Domain Event Memory.

• The event memory stores news under its assigned domain, retrieves similar units across domain memories, and combines potential-domain information with the explicit domain representation.Source: §3.3.2 · p. 5
• Per-domain event memories divide news into Q clusters, initialized from K-means cluster centers, with each memory unit representing similar news pieces.Source: §3.3.2 · p. 5

3.3.3 Discussion

• Unlike classifier-generated pseudo-labels, memory-bank label completion is trained end-to-end for fake-news detection and retains event information to improve transparency.Source: §3.3.3 · pp. 5–6

3.4 Domain Adapter

• Enriched domain information conditions the aggregation of useful cross-view representations for the final prediction.The Domain Adapter takes the Domain Memory Bank representation [cd, u] as input.Source: §3.4 · p. 6
• A feed-forward network produces weights that represent the importance of each cross-view representation.The weight vector can indicate which cross-view interactions a news piece's decision depends on.Source: §3.4 · p. 6

3.5 Predictor

• The Predictor uses the aggregated representation r to predict the probability that a news piece is fake.All parameters are optimized by back-propagation with cross-entropy loss.Source: §3.5 · p. 6

4 EXPERIMENTS

• The experiments ask whether M3FEND outperforms other approaches across different datasets.This is RQ1.Source: §4 · p. 6
• The experiments ask whether the M3FEND framework improves a real-world fake news detection system.This is RQ2.Source: §4 · p. 6
• The experiments examine the effects of the framework's different views and components.This is RQ3.Source: §4 · p. 6
• The experiments investigate how M3FEND models domain discrepancy, finds potential domain labels, and responds to hyperparameter changes.These are the remaining questions posed for the experiments.Source: §4 · p. 6

4.1.1 Datasets.

• The evaluation covers En-3, combining FakeNewsNet and COVID across three domains, and Chinese Ch-3, Ch-6, and Ch-9 datasets from Sina Weibo.Ch-3 has three domains; Ch-6 samples six daily-life domains from Ch-9, which contains nine domains.Source: §4.1.1 · p. 6
• The datasets are extended with emotion and style features, including high-level style measures such as readability, logic, credibility, and formality.The paper also uses low-level and high-level style features.Source: §4.1.1 · pp. 6–7

4.1.2 Baselines.

• Baselines span separately trained single-domain methods, mixed-domain methods, and designed multi-domain models, including EDDFN, MDFEND, MMoE, and MoSE.The baseline set also includes text encoders and methods that use style or emotion features.Source: §4.1.2 · pp. 7–8

4.1.3 Experimental details.

• Datasets are split into train, validation, and test sets in a 6:2:2 ratio, and reported Acc, macro F1, and AUC are averages over ten runs.The domain distribution is kept in each split; early stopping uses validation sets.Source: §4.1.3 · p. 8

4.2 Offline Results (RQ1)

• On most tasks, multi-domain methods outperform mixed-domain methods, while mixed-domain methods generally outperform separately trained single-domain approaches.The authors suggest well-designed sharing structures may alleviate domain conflict.Source: §4.2 · p. 8
• StyleLSTM and DualEmo outperform BiGRU, TextCNN, and RoBERTa, supporting the benefit of adding views to multi-domain detection.DualEmo is especially effective in Entertainment and Finance, outperforming most multi-domain methods there.Source: §4.2 · p. 8
• M3FEND significantly outperforms the best baseline in most tasks, improving overall and domain-specific detection performance.The reported significance is based on t-tests.Source: §4.2 · p. 8, Table 8 · p. 9
• MDFEND and M3FEND outperform EANN, MMoE, MoSE, and EDDFN on most tasks; the authors contrast their soft sharing with those models' hard sharing.The paper cites prior work reporting difficulty learning common knowledge with shared structures across many domains,.Source: §4.2 · p. 8
• M3FEND performs better than MDFEND on most tasks, while adding potential domain-label discovery and semantic, emotional, and stylistic views.MDFEND assumes the given domain label is exact and complete and extracts semantic information only.Source: §4.2 · pp. 8–9

4.3 Online Tests (RQ2)

• The online test set is highly skewed (roughly 300:1 real to fake), and training and test periods do not overlap.The test ran for one week using data collected by the deployed system.Source: §4.3 · p. 9
• SPAUCFPR≤0.1 is reported alongside AUC and F1 to assess detection under a low false-positive-rate constraint.The stated operational priority is improving true positive rate while minimizing false positives.Source: §4.3 · p. 9
• The authors report relative improvement over online RoBERTa and say M3FEND improves on the baselines across all metrics.They withhold online data details and absolute results due to company regulations.Source: §4.3 · p. 9

4.4 Ablation Study (RQ3)

• All three views benefit detection, with the semantic view especially important and the emotional view more effective than the stylistic view.The authors suggest emotion captures publisher and social characteristics, whereas style represents publisher preference.Source: §4.4 · p. 9
• The study compares view and component ablations across four datasets using overall F1 scores.Table 10 is identified as the results table for the ablation study.Source: §4.4 · p. 9
• Removing the interactor, memory, or adapter worsens results; the adapter is more effective than the interactor, and its memory inputs add useful implicit domain representations.Replacing the adapter with averaging also performs worse, supporting adaptive aggregation of discriminative representations.Source: §4.4 · pp. 9–10

4.5.1 Effectiveness of Domain Discrepancy Modeling

• Different domains yield different combinations of semantic, emotional, and stylistic cross-view representations, which the authors say facilitates modeling domain discrepancy.Figure 5 visualizes the importance of different views in a cross-view interaction.Source: §4.5.1 · p. 10
• Cross-view representation importance varies by domain; Entertainment values the emotion-heavy third cross-view most, consistent with its often-sensational news.Figure 6 compares the importance of four cross-view interactions across six domains.Source: §4.5.1 · p. 10

4.5.2 Effectiveness of Domain Label Completion

• A political-news case study’s predicted domain-label distribution indicates that Domain Event Memory captures potential domain-label distributions.The case study shows representative examples from the most related memory units.Source: §4.5.2 · p. 10

4.5.3 Hyperparameter Sensitivity

• On Ch-6, even the worst settings across four extractor-channel and interactor-head choices reach F1=0.9149, above MDFEND’s best F1:0.9093.Performance varies non-monotonically with β: large values risk overfitting recent samples, while small values risk underfitting them.Source: §4.5.3 · p. 10, Figure 7 · p. 11

Fake News Detection

• Prior fake-news methods are grouped into content-based and social-context-based approaches,; M3FEND uses semantic, style, and emotion features in the content-based group.Content-based approaches use news content and factual sources; social-context approaches use user engagements.Source: §5 · p. 11

Multi-domain Fake News Detection

• Unlike EDDFN’s hard sharing, M3FEND soft-shares domain-specific knowledge; unlike MDFEND, it uses Domain Memory Bank to address incomplete single-domain labels.M3FEND also explores knowledge sharing across emotional and stylistic views and adaptive cross-view knowledge.Source: §5 · p. 11

Multi-domain Learning

• M3FEND is framed as modeling domain relationships rather than learning domain-invariant representations.The passage reports that M3FEND outperforms existing methods.Source: §5 · p. 11

• M3FEND demonstrated superior effectiveness over other state-of-the-art approaches in offline experiments and superiority in online tests.Source: §6 · p. 11
• The Domain Adapter aggregates cross-view representations for prediction to address domain shift.Source: §6 · p. 11
• The Domain Memory Bank discovers potential domain labels and models domain characteristics to address domain labeling incompleteness.Source: §6 · p. 11
• The framework extracts news representations from multiple views and automatically models cross-view interactions.Source: §6 · p. 11