Diving Deep into Clickbaits: Who Use Them to What Extents in Which Topics with What Effects?

I. INTRODUCTION

• Clickbait headlines use linguistic techniques to entice clicks without delivering on their promises, while their use has expanded across the media industry.The introduction notes concerns about reader trust and brand value, alongside the possibility of temporary advertising revenue gains.Source: §I · p. 1
• The study asks how extensively mainstream and unreliable media use clickbait, whether topic distributions differ, and which headlines generate more engagement.The authors note that little is known about clickbait’s prevalence across media types or its engagement on social networks.Source: §I · p. 1
• The authors analyze 1.67 million Facebook posts from 68 mainstream and 85 unreliable media organizations using a supervised detector based on distributed subword embeddings.The model maps words to 300-dimensional embeddings and achieves 98.3% accuracy on a labeled dataset.Source: §I · pp. 1–2
• The paper examines clickbait from multiple perspectives, including qualitative and quantitative analyses of content and its impact.It also studies topic distributions and engagement through reactions, comments, and shares.Source: §I · pp. 1–2

II. DATASET

• The classifier is developed using 32,000 manually labeled headlines, with each headline labeled by at least three volunteers.The dataset contains 15,999 clickbait and 16,001 non-clickbait headlines.Source: §II · p. 2
• The Media Corpus contains 1.67 million link and video posts from 68 mainstream and 85 unreliable media organizations, collected from 2014 to 2016.The source collection exceeded two million posts; the analysis excludes photo posts and includes headlines, status messages, reactions, and some article bodies.Source: §II · p. 2

III. CLICKBAIT DETECTION

• The detection section frames the study’s first step as identifying clickbait and non-clickbait headlines; Figure 1 compares average daily posting by mainstream and unreliable media.Its green and blue areas indicate which media type posted more during each period.Source: §III · p. 2, Figure 1 · p. 2

A. Problem Definition

• Clickbait identification is a supervised binary classification task that maps labeled sentences to clickbait or non-clickbait classes.The paper describes learning a function from sentences to one of the two classes.Source: A. Problem Definition · p. 2

B. Problem Modeling

• Skip-Gramsw extends continuous skip-gram embeddings with subword information, using word substrings rather than bag-of-words features.The authors contrast this representation with prior use of bag-of-words and hand-crafted rules.Source: B. Problem Modeling · p. 2

1) Skip-Gramsw:

• Rather than treating each word as a unit, Skip-Gramsw predicts context subwords from a given subword, sharing representations across words to support rare words.For example, it predicts “qui” and “ick” in context given “uic” as input.Source: 1) Skip-Gramsw · p. 3
• The model learns subword vector representations by predicting nearby subwords in context.For a given subword, it aims to predict context subwords.Source: Figure 3 · p. 3, 1) Skip-Gramsw · p. 3
• A word embedding is the sum of the vector representations of its subwords.The weights to the hidden layer form the subword vector representations.Source: 1) Skip-Gramsw · p. 3

2) Pre-trained Vectors:

• Skip-Gramsw learns pre-trained word vectors from Media Corpus text, without requiring labeled data for embedding training.The corpus includes headlines, messages, and bodies; the paper compares these vectors with Google News vectors.Source: 2) Pre-trained Vectors · p. 3

3) Classification:

• For classification, the method averages a sentence’s word embeddings into a hidden representation and trains a linear classifier using softmax.The softmax computes a probability distribution over the classes.Source: 3) Classification · p. 3

C. Evaluation

• Evaluation used 10-fold cross-validation, averaged across five runs, on the Headlines dataset, with accuracy, precision, recall, f-measure, ROC-AUC, and Cohen’s κ as metrics.The paper notes that earlier studies used a smaller dataset whose training and test sets were unavailable for comparison on the same test bed.Source: C. Evaluation · p. 3
• 0.975 f-measure (2.5% higher than Chakroborty et al.) and a κ score of 0.952 were achieved by Skip-Gramsw without pre-trained vectors; the headlines, bodies, and messages vectors performed best overall.The authors report that differences among methods were small, but regard small improvements above 0.95 as significant.Source: C. Evaluation · pp. 3–4
• The Media Corpus has 477,236 unique embeddings, and its smaller corpus contributes more to clickbait classification than the 100-billion-embedding Google News dataset.The authors suggest that the Media Corpus embeddings contain more domain-specific knowledge.Source: C. Evaluation · p. 4
• The paper proceeds to use the best-performing clickbait classifier for a large-scale study of clickbait practice on Facebook.The cited model is Skip-Gramsw with pre-trained vectors from headlines, bodies, and messages.Source: C. Evaluation · p. 4

IV. PRACTICE OF USING CLICKBAIT IN SOCIAL NETWORK

• The study analyzes clickbait practice on Facebook using the Media Corpus from three perspectives.Source: §IV · p. 4

A. Quantitative Analysis

• 39.26% of unreliable-media headlines and 33.54% of mainstream-media headlines were clickbait.The counts were 308,095 of 784,665 for unreliable media and 297,774 of 887,794 for mainstream media.Source: A. Quantitative Analysis · p. 4
• Clickbait-to-non-clickbait ratios were 61.64% in non-news broadcast media versus 22.32% in news-oriented broadcast media, whose level was described as close to print media.Satire led link and video clickbait, while print had the lowest link clickbait and conspiracy had the lowest video clickbait among media categories.Source: A. Quantitative Analysis · p. 4
• 61.64% of non-news broadcast headlines were clickbait, versus 22.32% for news-oriented broadcast media.News-oriented broadcast media had about 25% clickbait at its distribution peak, while non-news broadcast media peaked near 60%.Source: Figure 5 · p. 4, A. Quantitative Analysis · p. 4

B. Qualitative

• Biterm Topic Modeling identifies topics by modeling aggregated word co-occurrence patterns in short documents.The authors chose it because traditional document-level topic models may struggle with sparse word co-occurrences in short headlines.Source: B. Qualitative · pp. 4–5
• Clickbait headlines in print and broadcast skew toward personalized, sensationalized, and entertaining topics, unlike non-clickbait headlines’ focus on public policies and civic affairs.The topic model presents five topics for clickbait and non-clickbait content in each media category, each represented by ten words.Source: B. Qualitative · p. 5
• Four of five print-media clickbait topics concern Donald Trump’s views on women, with the words “Trump,” “woman,” “make,” and “new” in each.The authors’ manual search found question-based and other clickbait headlines in stories about Trump and women.Source: B. Qualitative · p. 5
• Broadcast clickbait topics mostly concern entertainment and lifestyle, while unreliable-media clickbait topics range from politics to lifestyle.At least three unreliable-media topics appeared political; non-clickbait topics across media primarily concern law and order and the U.S. presidential campaign, covering 12 of 15 topics.Source: B. Qualitative · p. 5

C. Impact

• Clickbait posts generally receive more attention and reach more users than non-clickbait posts, except in broadcast media.Attention is measured using reactions, comments, and shares; the figure’s blue and green areas mark which post type receives more average attention.Source: C. Impact · pp. 5–6
• Conspiracy media organizations repost the same links more than other categories for both clickbait and non-clickbait, and clickbait media repeat clickbait links more.Figure 6 includes links reposted at least 20 times; bar height represents how many times a link was posted on Facebook by a media category.Source: C. Impact · p. 6
• Media also use clickbait status messages to promote non-clickbait articles, and the clickbait media category leads this practice.Source: C. Impact · p. 6, Table VI · p. 6

V. RELATED WORK

• Clickbait’s sensational language and entertaining story focus have roots in tabloid press, while the Internet and social media enable faster, broader publication and reach,.Source: §V · p. 6
• The authors found no prior research measuring clickbait’s extent across mainstream and alternative web media or testing whether it increases social-media engagement.Earlier studies examined headline attributes and engagement correlations, but the authors identify these two questions as unresolved.Source: §V · p. 7
• Earlier studies included a 151-article manual analysis, a 2,057-adult experiment, and headline testing by Chartbeat and HubSpot and Outbrain.The adult experiment found more negative reactions to clickbait political headlines, while HubSpot and Outbrain suggested clickbait may increase temporary engagement.Source: §V · p. 7
• Prior detection methods used hand-crafted features or headline-only signals, which the authors argue can be domain-specific or fail to assess whether a headline fairly represents its article, [10],,.Other work generated clickbait headlines with an RNN trained on two million headlines.Source: §V · pp. 7–8

• The word-embedding model performs better than embeddings based on the Google News dataset.Source: §VI · p. 8
• Mainstream media are increasingly involved in clickbait practice.Source: §VI · p. 8
• Broadcast media use clickbait at a higher percentage than print media, with non-news broadcast media mostly contributing to that practice.Source: §VI · p. 8
• Unreliable media use clickbait at a higher percentage than other media.Source: §VI · p. 8
• Topic modeling indicates that clickbait is prevalent in personalized and entertaining areas.Source: §VI · p. 8
• Future work will incorporate news content when defining the clickbaitiness of a headline.Source: §VI · p. 8