Source-linked AI summary
Event detection, tracking, and visualization in Twitter: a mention-anomaly-based approach
Adrien Guille, Cecile Favre
TL;DR
Twitter’s large stream of unrelated content makes significant events difficult to detect, and existing methods often overlook its social connections. MABED detects events using mention activity and estimates each event’s discussion period dynamically. Across English and French data, the authors report better performance than a mention-ignoring variant and two recent methods, with improved robustness to noisy content.
Problem
Twitter’s high volume of unrelated tweets makes significant events difficult to identify, while methods focused on text overlook the social aspect of the platform.
Source: §1 · pp. 1–2Method
MABED uses tweet-based statistical measures, including mention creation frequency, to detect events and estimate each event’s discussion period dynamically.
Source: §1 · p. 2, §3.2 · p. 4Results
MABED outperformed α-MABED and two recent literature methods in the reported tests, and the authors report improved robustness to noisy Twitter content.
Source: §1 · p. 2, §6 · p. 15Takeaways & Limitations
The authors report that MABED can help interpret detected events, understand users’ interests, and explore events through three visualizations.
Source: §1 · p. 2, §6 · p. 16The method approximates the mention-containing tweet count distribution with a normal distribution when the count is large.
Source: §3.3 · p. 5
Abstract
from arXiv · showhide
The ever-growing number of people using Twitter makes it a valuable source of timely information. However, detecting events in Twitter is a difficult task, because tweets that report interesting events are overwhelmed by a large volume of tweets on unrelated topics. Existing methods focus on the textual content of tweets and ignore the social aspect of Twitter. In this paper we propose MABED (i.e. mention-anomaly-based event detection), a novel statistical method that relies solely on tweets and leverages the creation frequency of dynamic links (i.e. mentions) that users insert in tweets to detect significant events and estimate the magnitude of their impact over the crowd. MABED also differs from the literature in that it dynamically estimates the period of time during which each event is discussed, rather than assuming a predefined fixed duration for all events. The experiments we conducted on both English and French Twitter data show that the mention-anomaly-based approach leads to more accurate event detection and improved robustness in presence of noisy Twitter content. Qualitatively speaking, we find that MABED helps with the interpretation of detected events by providing clear textual descriptions and precise temporal descriptions. We also show how MABED can help understanding users' interest. Furthermore, we describe three visualizations designed to favor an efficient exploration of the detected events.
1 Introduction
Twitter offers timely information but its volume makes significant events difficult to identify, while keyword trends omit crowd attention and timing. MABED uses mention activity and dynamically estimated event periods, and its evaluation reports stronger detection performance and robustness to noisy content.
1 Introduction
- Twitter’s information overload makes it difficult to identify significant events, while keyword trends omit crowd-attention levels and temporal indications.The introduction frames automated significant-event detection and tracking as useful for analyzing which events or event types interest the crowd.Source: §1 · p. 1
- Prior event-detection work largely emphasizes tweet text and neglects Twitter’s social links, including time- and topic-specific mentions between users.The introduction describes mentions as links created when users engage others, reply, or retweet.Source: §1 · pp. 1–2
- MABED uses tweet-based statistical measures and mention frequency to produce events with a main word, weighted related words, a time period, and impact magnitude.Unlike fixed-duration approaches, it dynamically estimates how long each event is discussed; the paper also introduces three interactive visualizations for exploring events.Source: §1 · p. 2
- MABED outperformed α-MABED and two recent literature methods in the reported tests, while mention frequency was associated with more accurate detection and robustness to noisy content.The evaluation used English and French corpora of about 1.5 million and 2 million tweets, respectively, and also found short computation times.Source: §1 · p. 2
- The paper also analyzes detected event types alongside author communities in the following network, using the results to examine the relationship between social and topical structures.The authors report that this analysis can help understand users’ interests.Source: §1 · p. 2
2 Related Work
Prior work detects events in tweet streams through textual bursts, topic evolution, and clustering, while visualization tools present event descriptions and temporal activity. The reviewed approaches also raise concerns about short-text modeling, scalability, external-data influence, and merging concurrent events.
2.1 Event Detection and Tracking from Tweets
- Peakiness Score identifies words specific to fixed-length windows, while Trending Score ranks n-grams by normalized frequency within and across windows.The review notes that individual words can lack the context needed to describe complex events.Source: §2.1 · p. 3
- Topic models track evolving word distributions or align expanded tweets with media event transcripts, but the review cites scalability and parallel-event handling concerns.ET-LDA’s use of search-engine expansion and traditional-media retranscriptions heavily influences its results.Source: §2.1 · p. 3
- EDCoW clusters words with correlated bursts, while TwEvent and ET cluster similar n-grams or bigrams within fixed-length windows.The review notes that cross-correlation can be costly and can cluster unrelated simultaneous events; TwEvent also relies on Wikipedia and Microsoft Web N-Gram data.Source: §2.1 · p. 3
2.2 Event Visualization in Twitter
- Tweet-visualization systems display event word sets or clouds, related-tweet volume over time, timelines, and heatmaps of event distributions.TwitInfo lets users inspect tweets from an event’s time interval or the most-cited URLs in those tweets.Source: §2.2 · p. 3
3 Proposed Method
MABED seeks the k events with the greatest impact on tweeting behavior, representing each as a bursty topic with a time interval and impact score. Its three components detect mention anomalies, select descriptive words, and return a list that merges duplicates.
3.1 Problem Formulation
- A bursty topic is a main term and weighted descriptive words that attract unusually high mention-creation attention during an interval; an event adds that interval and an impact score.The method’s output is a list of k events with the highest impact on crowd tweeting behavior.Source: §3.1 · p. 4
3.2 Overview of the Proposed Method
- MABED first ranks partially described events by impact, then selects related words using co-occurrence and temporal dynamics, and finally merges duplicates into the top-k list.The event-detection stage uses mention-creation frequency for each word in the mention-containing vocabulary.Source: §3.2 · pp. 4–5
3.3 Detection of Events Based on Mention Anomaly
- For each word, MABED finds the contiguous interval with the greatest summed mention-frequency anomaly, allowing negative slices when they increase total impact.This avoids splitting multi-day events during periods of lower Twitter activity and permits each word to serve as the main word of at most one event.Source: §3.3 · pp. 5–6
3.4 Selection of Words Describing Events
- MABED selects candidate descriptive words by co-occurrence with the main word during the event interval, then retains words whose weights meet threshold θ.The fixed candidate limit p bounds description length; parameters p and θ let users adjust the level of detail.Source: §3.4 · p. 6
3.5 Generating the List of the Top k Events
- The event graph stores event descriptions, while the redundancy graph records links between main words of duplicate events; duplicates are merged once k distinct events have been processed.The resulting list contains the top k most impactful distinct events.Source: §3.5 · pp. 6–7
4 Experiments
Experiments on English and French Twitter corpora compare MABED with baselines and evaluate its event descriptions, timing, and community-interest analysis. The results report strong detection performance alongside benefits from mention-aware detection and dynamically estimated event durations.
4.1 Experimental Setup
- Evaluation was limited to the 40 most impactful events per method in each corpus; two annotators independently rated meaningfulness and significance, tracked duplicates, and the study measured precision, recall, and duplicate-event rate.Source: §4.1 · p. 9
4.2 Quantitative Evaluation
- MABED gained an average relative 17.2% in F-measure over α-MABED, which ignores mentions, supporting the relevance of mention behavior to event detection.Source: §4.2 · p. 9
- MABED outperformed baselines in F-measure by a bigger margin on Cen, where up to 50% of tweets were non-event-related, suggesting greater robustness to noisy content.The authors also report that most non-significant events detected by α-MABED, TS2, and TS3 were spam-related and were not detected by MABED.Source: §4.2 · p. 9
- MABED achieved the best performance on both corpora: precision 0.775 and F-measure 0.682 on Cen, and precision and F-measure 0.825 on Cfr.The comparison reports performance for five methods across both corpora.Source: §4.2 · p. 9, Table 3 · p. 10
- Runtime grew linearly with subsample size, and MABED with 8 threads was on average 67% faster than MABED.Reported runtimes exclude vocabulary preparation and term-frequency precomputation.Source: Figure 5 · p. 9, §4.2 · p. 10
- On Cen, the overlap threshold σ mainly affects DERate, which increases as σ rises; the highest F-measure occurs for σ from 0.2 to 0.5.Even at σ = 1, MABED’s F-measure of 0.582 exceeded all Cen baselines.Source: Figure 6 · p. 9, §4.2 · p. 10
4.3 Qualitative Evaluation
- MABED’s ranked, bounded event descriptions aid interpretation, while dynamically estimated intervals provide finer temporal precision than the daily reporting typical of existing methods.For the Fort Hood shooting, MABED first detected activity in the 19:30–20:00 UTC time-slice; the event description spans November 5 to November 8.Source: §4.3 · pp. 10–12
4.4 Analysis of Detected Events
- −0.36 correlation between c0 and c1 category-weight distributions contrasts with 0.93 between Cen and Cen(random), suggesting the detected communities differ in interests.The authors conclude that detected communities are relevant to users’ interests.Source: §4.4 · pp. 12–13, Figure 9 · p. 13
5 Implementation and Visualizations
MABED’s visualizations support chronological browsing, inspection of event-impact patterns, and exploration of topical relationships. A year-long French political Twitter deployment used them to monitor daily events and examine how attention shifted over time.
Time-oriented visualization
- The time-oriented view places events on a chronological ribbon and shows the selected event’s MABED description, image, and hypertext.The hypertext and image come from the top Bing result, queried with the MABED description.Source: §5 · p. 13, Figure 10 · p. 14
Impact-oriented visualization
- The impact-oriented chart plots each event’s mention-anomaly function, revealing single, diminishing, or increasing peaks in user reactions.Users can filter the functions with the interactive legend.Source: §5 · pp. 13, 15, Figure 11 · p. 14
Topic-oriented visualization
- The topic graph encodes events as grey main-term nodes and related words as blue nodes, helping identify similar events and words shared across events.In the example, two presidential candidates appear across many events but together in only one.Source: §5 · p. 15
Case study: monitoring the French political conversation
- From December 2013 to November 2014, MABED analyzed French political tweets in real time and refreshed visualizations every 10 minutes with the five most impactful events from the prior 24 hours.The highest reported crawling rate was about 150,000 tweets in 24 hours.Source: §5.1 · p. 15
- The two monitored days showed successive attention to related events on December 13 and simultaneous discussion of loosely related events on December 14.The authors suggest related events may compete or collaborate, with competing events negatively correlated and cooperating events positively correlated in mention-anomaly patterns.Source: §5.1 · p. 15
6 Conclusion
MABED detects Twitter events using mention frequency and dynamically estimates their discussion periods. The authors report better performance than a mentions-free variant and two recent methods, and describe applications to community-interest analysis, event visualization, and French political monitoring.
Conclusion
- MABED uses mention-creation frequency as a social signal and dynamically estimates how long each event is discussed, limiting event fragmentation.The method has linear runtime in corpus size.Source: §6 · p. 15
- MABED outperformed α-MABED, which ignores mentions, and two recent literature methods in all reported tests.The authors also report improved robustness with noisy Twitter content.Source: §6 · p. 15
- Analysis of detected events alongside user communities helped reveal the interplay between Twitter’s social and topical structure and users’ community interests.The conclusion states that MABED can help understand user communities’ interests.Source: §6 · pp. 15–16
- The authors presented three visualizations for exploring detected events and used MABED with them to monitor French political conversation from December 2013 through November 2014.They also report that highlighting main words improves event-description readability and that MABED’s temporal information helps indicate when real-world events occurred.Source: §6 · pp. 15–16
- Future work includes adding discussion features such as distinct-user counts and geolocations, and incorporating sentiment analysis to enrich event descriptions.Source: §6 · p. 16