Source-linked AI summary
XFake: Explainable Fake News Detector with Visualizations
Fan Yang, Shiva K. Pentyala, Sina Mohseni, Mengnan Du, Hao Yuan, Rhema Linder, Eric D. Ragan, Shuiwang Ji, Xia Hu
TL;DR
Fake-news fakeness can stem from attributes or context beyond a statement’s text, while users need explanations for model predictions. XFake combines attribute, semantic, and linguistic analysis with supporting examples and visualizations. It provides fake-news prediction scores alongside explanations to support users’ credibility decisions.
Problem
Fake-news fakeness may depend on attributes such as context or speaker rather than statement content, and users need explanations to assess predictions.
Source: §1 · p. 1Method
XFake combines attribute, semantic, and linguistic analysis frameworks and provides explanations, supporting examples, and visualizations.
Source: §1 · p. 1Results
XFake provides fake-news prediction scores with explanations from multiple perspectives, supporting examples, and visualizations.
Source: §1 · p. 1Takeaways & Limitations
XFake is designed to help users identify news credibility through predictions and explanations.
Source: §4 · p. 4
Abstract
from arXiv · showhide
In this demo paper, we present the XFake system, an explainable fake news detector that assists end-users to identify news credibility. To effectively detect and interpret the fakeness of news items, we jointly consider both attributes (e.g., speaker) and statements. Specifically, MIMIC, ATTN and PERT frameworks are designed, where MIMIC is built for attribute analysis, ATTN is for statement semantic analysis and PERT is for statement linguistic analysis. Beyond the explanations extracted from the designed frameworks, relevant supporting examples as well as visualization are further provided to facilitate the interpretation. Our implemented system is demonstrated on a real-world dataset crawled from PolitiFact, where thousands of verified political news have been collected.
1 INTRODUCTION
Fake news detection is challenging because fakeness may arise from attributes or context beyond a news item’s text, and users need explanations to assess predictions. XFake addresses these challenges with multiple analysis frameworks and supporting explanations.
1 INTRODUCTION
- The spread of fake news accompanies easier public access to news through websites and social media.The paper describes fake news as potentially misleading the public with false or biased information [10].Source: §1 · p. 1
- Effective detection is challenging because fakeness can depend on context or speaker rather than news content, and users need explanations to inform decisions.The source of fakeness is described as instance-dependent.Source: §1 · p. 1
- XFake analyzes news attributes, statement semantics, and statement linguistics, and supplies explanations, supporting examples, and visualizations.The system is described as helping users understand why a news item receives a particular prediction.Source: §1 · p. 1
- To the authors’ knowledge, XFake is the first explainable fake-news detector that analyzes news from multiple perspectives.Source: §1 · p. 1
2 XFAKE ARCHITECTURE
XFake combines attribute, semantic, and linguistic analysis frameworks trained on verified news. It produces a weighted fake-news score and explanations through feature importance, word or phrase attribution, supporting examples, and visualizations.
2 XFAKE ARCHITECTURE
- XFake takes target news and its attributes as input, outputs predictions and explanations, and is trained on a verified news set.Source: §2 · p. 2
- MIMIC uses a neural-network teacher’s soft labels to train an 80-tree student, whose attribute importance scores support explanations.The approach aims to combine neural-network performance with the explainability of tree ensembles.Source: §2 · p. 2
- ATTN combines pretrained embeddings, convolution, and self-attention; attention weights support instance-level explanations, while kernel sizes yield one-, two-, or three-gram analyses.The passage specifies word2vec embeddings with 300 dimensions and 512-dimensional representations at each spatial location.Source: §2 · p. 2
- PERT uses eight linguistic features with an XGBoost classifier and estimates feature significance by replacing a feature with same-distribution random noise and checking prediction differences.The listed features include part-of-speech ratios, sentiment, normalized text length, and question-mark or exclamation-mark presence.Source: §2 · p. 2
- XFake combines MIMIC, ATTN, and PERT outputs as a weighted sum tuned on validation-set performance, producing a score that indicates the probability news is fake.MIMIC explains key news components, ATTN attributes words or phrases, and PERT explains linguistic features; supporting examples are retrieved using MIMIC- or ATTN-derived explanations.Source: §2 · p. 2
- Supporting examples are retrieved from training samples using important attributes or words and receive similarity scores based on their match to the input.Higher-scored examples are considered more informative for delivering explanations.Source: §2 · p. 3
- Visualizations use histograms for numerical values, heatmaps for word or phrase importance, and interactive tree diagrams for overall structure and activated paths.Source: §2 · p. 3
3 SYSTEM DEMONSTRATION
XFake is demonstrated on PolitiFact news, with a human study reporting a trade-off between users’ prediction speed and accuracy. A use case shows its prediction and multi-perspective explanations, alongside supporting examples and interactive decision trees.
3 SYSTEM DEMONSTRATION
- XFake is implemented in Python and deployed with a Flask backend and HTML front end.The demonstration introduces the data source before showing a specific use case.Source: §3 · p. 3
3.1 Data Source
- The training, validation, and test data come from PolitiFact and contain 5,104 political news items, split into 4,083, 510, and 511 items.The original labels are transformed into binary True and False labels; the test set is used for overall evaluation.Source: §3.1 · p. 3
3.2 Use Cases
- On the validation set, MIMIC, ATTN, and PERT achieve accuracies of 67.1%, 67.3%, and 53.2%, respectively.Their weighted-sum coefficients are 0.36, 0.36, and 0.28, tuned according to framework performance.Source: §3.2 · p. 3
- Explanations helped users better understand and predict system behavior, but required more time to review and interpret detection results.The study involved 147 valid AMT users and measured prediction accuracy and time.Source: §3.2 · p. 3
- Figure 6 pairs supporting news matched through MIMIC attributes or an ATTN word with visualized trees showing each tree’s activated decision path.MIMIC examples include similarity scores; users can expand or compress the 80 interactive trees.Source: §3.2 · p. 3, Figure 6 · p. 4
4 CONCLUSIONS AND FUTURE WORK
XFake jointly analyzes news from multiple perspectives and provides explanations through attributes, word and phrase significance, linguistic features, and supporting examples. Its visualizations are intended to facilitate users’ decision making; future extensions could cover multi-source or multimodal news.
4 CONCLUSIONS AND FUTURE WORK
- XFake jointly analyzes news from multiple perspectives and provides explanations through attribute importance, word and phrase significance, linguistic features, and supporting examples.The conclusion describes the system as designed to help end-users identify news credibility.Source: §4 · p. 4
- The authors state that XFake’s visualizations further enhance explainability and facilitate users’ decision making.Source: §4 · p. 4
- The authors identify multi-source or multimodal news data as promising future extensions beyond PolitiFact and textual news.Source: §4 · p. 4