PaperlayerView exact revision on arXiv

Source-linked AI summary

Multiscale Event Detection in Social Media

Xiaowen Dong, Dimitrios Mavroeidis, Francesco Calabrese, Pascal Frossard

arXiv:1404.7048v2cs.SIcs.LGphysics.soc-phstat.ML

TL;DR

Many event-detection methods use fixed temporal and spatial resolutions, although social-media events occur across different scales. This paper uses wavelets and graph clustering to handle those scales, and reports effective detection on synthetic and Twitter data.

  • Problem

    Many event-detection approaches rely on fixed temporal and spatial resolutions, although events span different intervals in time and space and the interaction between these dimensions is unclear.

    Source: §1 · p. 2

  • Method

    The approach uses wavelet properties to handle temporal–spatial scale interactions, clusters a similarity graph built at appropriate scales, and uses spatiotemporal noise analysis to define term filtering.

    Source: §1 · p. 3

  • Results

    Experiments on synthetic and Twitter data show the proposed approach can effectively detect events of different temporal and spatial scales.

    Source: §1 · p. 3

  • Takeaways & Limitations

    Modeling temporal–spatial scale relationships helps separate events of different scales simultaneously and meaningfully, and the framework can generalize to other multiscale analysis domains.

    Source: §9 · p. 30

    More aggressive filtering can reduce noise and computational cost, but may also remove terms related to relatively small-scale events.

    Source: §7.4 · p. 28

Abstract

from arXiv · show

Event detection has been one of the most important research topics in social media analysis. Most of the traditional approaches detect events based on fixed temporal and spatial resolutions, while in reality events of different scales usually occur simultaneously, namely, they span different intervals in time and space. In this paper, we propose a novel approach towards multiscale event detection using social media data, which takes into account different temporal and spatial scales of events in the data. Specifically, we explore the properties of the wavelet transform, which is a well-developed multiscale transform in signal processing, to enable automatic handling of the interaction between temporal and spatial scales. We then propose a novel algorithm to compute a data similarity graph at appropriate scales and detect events of different scales simultaneously by a single graph-based clustering process. Furthermore, we present spatiotemporal statistical analysis of the noisy information present in the data stream, which allows us to define a novel term-filtering procedure for the proposed event detection algorithm and helps us study its behavior using simulated noisy data. Experimental results on both synthetically generated data and real world data collected from Twitter demonstrate the meaningfulness and effectiveness of the proposed approach. Our framework further extends to numerous application domains that involve multiscale and multiresolution data analysis.

1 Introduction

Social media offers rapid, large-scale event information, but events span different temporal and spatial scales while many detection methods use fixed resolutions. The paper proposes wavelet-based multiscale detection and evaluates it on synthetic and Twitter data.

1 Introduction

  • Social media's real-time nature and large user base provide faster awareness of real-world happenings and broad coverage from different perspectives.Source: §1 · p. 2
  • Fixed temporal and spatial resolutions tend to produce detected events of similar scales, while how the two dimensions interact across resolutions remains unclear.Social media streams also contain noisy information irrelevant to events of interest.Source: §1 · pp. 2–3
  • The proposed method uses wavelet properties to handle interactions between temporal and spatial scales and clusters a similarity graph computed at appropriate scales.The paper also analyzes spatiotemporal noise distributions to define a term-filtering procedure and study behavior on simulated noisy data.Source: §1 · p. 3
  • Experiments on synthetic and Twitter data show the proposed approach can effectively detect events of different temporal and spatial scales.The authors compare it with their baseline approach.Source: §1 · p. 3

2 Spatiotemporal detection of events

The paper frames social-media event detection as identifying real-world happenings reflected in data concentrated in time, space, or both. Its objective is to detect events across spatiotemporal scales while remaining robust to ambiguous and noisy information.

2 Spatiotemporal detection of events

  • Events are real-world happenings reflected in social-media data concentrated in both time and space, or in at least one dimension.Source: §2 · p. 4
  • Tweets unrelated to ongoing events can bury event-relevant tweets when they make up a large part of the input.Source: §2 · p. 4
  • The paper's objective is to detect events spanning multiple temporal and spatial intervals while resisting ambiguous and noisy information in Twitter streams.Source: §2 · p. 4
  • The paper casts detection as graph-based clustering: tweets are vertices, similarity edges reflect temporal, spatial, and textual features, and clusters represent events.It compares a spatiotemporal-constraint baseline with a wavelet-based similarity scheme and studies noise using a homogeneous Poisson process.Source: §2 · p. 4

3 Local event detection via spatiotemporal constraints

The local event detection baseline builds a tweet-similarity graph using strict temporal and spatial thresholds, then clusters it and filters clusters into local events. Its thresholds suit events of similar scales but can split larger events or admit more noise when scales differ.

3 Local event detection via spatiotemporal constraints

  • The baseline sets tweet similarity to tf-idf text similarity only when both temporal and spatial differences meet thresholds; otherwise similarity is zero.Temporal difference is measured in minutes and spatial distance in meters.Source: §3 · p. 5
  • The method places tweets in an undirected weighted graph and applies the non-recursive Louvain method, which clusters by modularity without requiring a preset cluster count.Source: §3 · p. 5
  • A cluster counts as a local event only when its tweet count and number of distinct users exceed set thresholds.Source: §3 · p. 6
  • Thresholds tuned to expected event spans can detect similar-scale events concentrated in both dimensions, but low thresholds may split events and high thresholds may admit more noise.Source: §3 · p. 6
  • Further recursive grouping may not clearly preserve the resulting events' spatiotemporal-scale interpretation, and deciding when to stop can be difficult.Source: §3 · p. 6

4 Multiscale event detection using wavelets

The approach models the interaction between temporal and spatial scales, then uses wavelet-based similarities to support multiscale event detection through graph clustering.

4.1 Relationship model between temporal and spatial scales

  • Tweets can be considered similar when they match at a fine resolution in at least one dimension, trading off temporal and spatial scale rather than requiring both.The model allows coarse temporal resolution when tweets share terms and are close in space, and coarse spatial resolution when they are close in time.Source: §4.1 · p. 7
  • Comparing common-term occurrence time series avoids simply relaxing distance thresholds, which risks text ambiguity, or using exact distances, which might let one scale dominate.The series are constructed for the two tweets’ geographical cells using initial temporal and spatial resolutions.Source: §4.1 · p. 8

4.2 Wavelet-based similarity computation

  • Haar DWT approximation coefficients aggregate time series into coarser scales by factors of two, so correlation between coefficients at a selected level measures similarity at that temporal scale.The temporal comparison scale is selected inversely to the spatial scale: distant cells use finer temporal comparisons, while nearby cells can use coarser ones.Source: §4.2 · p. 9, Figure 4 · p. 10
  • MED combines text similarity with the maximum similarity among shared-term time series, constructs a weighted graph, and applies Louvain clustering to identify event clusters.Unlike LED’s fixed temporal and spatial constraints, MED compares common-term spatiotemporal patterns to accommodate events of different scales.Source: §4.2 · pp. 10–11
  • The initial resolutions need not match the true event scales; the paper suggests choosing nscale no higher than ⌈min(log2ld, log2lt)⌉.The bound reflects the number of spatial cells and the keyword time-series length, which constrain meaningful spatial scales and Haar DWT levels.Source: §4.2 · p. 11

5 Spatiotemporal analysis of noise in Twitter

The section characterizes common-term tweets as noise through spatial and temporal statistical analysis, then uses the analysis to motivate K-based term filtering and simulated noisy experiments.

5.1 Spatial distribution of noise in Twitter data

  • The spatial analysis tests common-term tweet locations for complete spatial randomness and uses Ripley’s K-function to assess proximity to a homogeneous Poisson process.Terms meeting the randomness criterion may create similarity-graph edges treated as noise.Source: §5.1 · pp. 13–14
  • Frequent-term tweets in Manhattan were close to a homogeneous Poisson pattern, with only slightly higher concentrations at small distances.This held for the four examined terms and the larger collection of frequent terms across roughly 500 days.Source: §5.1 · pp. 14–16

5.2 Temporal distribution of noise in Twitter data

  • For frequent terms in a six-hour window, uniform timestamps could be rejected at the 5% level in only 27% of cases, suggesting weak temporal patterns for many terms.Source: §5.2 · p. 16

5 Spatiotemporal analysis of noise in Twitter

  • The analysis motivates a term filter based on low standardized K-function values and synthetic experiments with simulated noisy data.These uses are intended to help evaluate detection behavior across space-time settings.Source: §5.2 · p. 16

6 Synthetic experiments

Synthetic experiments show that MED handles events at differing temporal and spatial scales more robustly than LED, especially when events are not localized in both dimensions. The comparisons use generated event and noise tweets and examine parameter sensitivity across several settings.

6.1 Experimental setup

  • The synthetic setup generates tweets with event-relevant and noise terms, spatial and temporal event locations, and Poisson-distributed noise tweets.Event-relevant tweet counts per event are sampled from 3 to 10, with one of 59 signal terms in each such tweet.Source: §6 · p. 17, §6.1 · p. 17

6.2.1 Events concentrated in both time and space without noise

  • For 20 noise-free events localized in both time and space, MED performs better than LED at small parameter values and is less sensitive to parameter choice.LED improves as its thresholds approach or exceed the events’ true scales, then drops slightly when thresholds become too large.Source: §6.2.1 · p. 18, Figure 8 · p. 19

6.2.2 Events concentrated in only one dimension without noise

  • For noise-free events localized in only one dimension, MED maintains performance comparable to the both-dimensions case, while LED drops significantly.LED performs reasonably only when thresholds cover all event scales.Source: §6.2.2 · pp. 18–19, Figure 9 · p. 19

6.2.3 Events concentrated in both time and space with noise

  • With noise and events concentrated in both dimensions, MED outperforms LED over many parameter values and has more stable NMI and F-measure performance.For both methods, F-measure first rises as false negatives decrease, then falls at larger settings as false positives increase.Source: §6.2.3 · p. 20, Figure 10 · p. 21

6.2.4 Events concentrated in only one dimension with noise

  • With noise and events localized in only one dimension, both methods’ F-measure drops significantly, but MED has higher peak performance and greater stability than LED.Source: §6.2.4 · p. 20, Figure 11 · p. 21

6.2.5 Influence of parameter settings

  • At fixed Poisson parameter and 10 by 10 area, increasing the interval from 32 to 128 improves both algorithms; at interval 32, expanding area to 16 by 16 lowers performance as noise tweets increase quadratically.The temporal-interval gain is more dramatic for LED, especially at large parameter values.Source: §6.2.5 · p. 21

6 Synthetic experiments

  • Across the synthetic experiments, MED generally detects events of different scales better than LED and is less sensitive to parameter selection.LED performs well when events share both temporal and spatial localization and its thresholds are chosen near their scales.Source: §6.2.5 · p. 22

7 Real world experiments

The real-world experiments compare LED and MED on New York Twitter data, illustrating how their different approaches handle events across temporal and spatial scales and examining graph-construction scalability.

7.1 Data description

  • From November 2011 to March 2013, researchers collected New York-area tweets via Twitter’s Streaming API; after filtering untagged and out-of-bounds tweets, they had 16,449,769 geotagged tweets, before clear-location filtering.They subsequently removed tweets containing clear location indicators, which they did not consider events of interest.Source: §7.1 · pp. 22–23

7.2 Implementation details

  • Both methods run daily; LED uses 30-minute and roughly 100-meter locality thresholds, while MED filters terms and analyzes time series across four spatial scales.The implementation also removes small clusters, clusters dominated by one user, and clusters identified as job advertisements or traffic alerts.Source: §7.2 · p. 23

7.3 Event detection results

  • LED finds meaningful, highly localized clusters, but its strict temporal and spatial constraints can split tweets about the same event into separate clusters.On 2011-11-17, it detected 41 local clusters in middle and lower Manhattan.Source: §7.3 · pp. 23–24, Figure 13 · p. 25, Table 2 · p. 26
  • MED detects events spanning long time intervals or broad areas, including OWS protests and events concentrated in time but spread in space; LED misses some such clusters.The examples include the Raise Cache event and the Mastercard promotion, which LED misses because their tweets are not concentrated in both time and space.Source: §7.3 · pp. 24–25, Table 3 · p. 26, Figure 14 · p. 27

7.4 Scalability

  • Similarity-graph construction is O(n × avg connectivity), and MED’s term filtering reduced average connectivity by more than 40% relative to LED in the experiment.Average connectivity was 2% of the tweet count in the real-world experiment; more aggressive filtering may also remove terms associated with small-scale events.Source: §7.4 · pp. 27–28
  • For roughly 8,000 daily geotagged tweets and 36,000 terms, graph construction took a few seconds for LED and about five minutes for MED on the lab server.The paper attributes MED’s main computational cost to DWT computations.Source: §7.4 · p. 28

8 Related work

Related work covers event detection using social media’s temporal, spatial, and textual signals, but the authors identify explicit handling of interacting temporal and spatial scales as a gap.

8 Related work

  • Social-media event detection has used keyword burstiness and combinations of textual, temporal, and spatial signals, including wavelets and spatiotemporal clustering.The surveyed work spans Twitter and Flickr applications.Source: §8 · pp. 28–29
  • Prior scale-space approaches generally select a suitable scale, while wavelet-based clustering methods produce clustering solutions at multiple resolutions.The authors instead use wavelets to select resolutions for a single similarity graph.Source: §8 · p. 29
  • The authors identify limited explicit treatment of multiple spatiotemporal scales and their interaction, and present their approach as an attempt to model that relationship.They also describe their statistical analysis of noisy Twitter information as a first of its kind.Source: §8 · pp. 29–30

9 Conclusion

The conclusion presents a multiscale event-detection approach that models relationships between temporal and spatial scales, alongside statistical analysis of noisy Twitter information. It identifies improved scale modeling, noise analysis, and scalability as directions for future work.

9 Conclusion

  • The proposed approach models relationships between temporal and spatial scales to separate events of different scales simultaneously and meaningfully.The conclusion frames this relationship as important to multiscale event detection.Source: §9 · p. 30
  • Statistical modeling of noise in Twitter streams informs a term-filtering procedure and offers insight into noise’s influence on event-detection algorithm design.Source: §9 · p. 30
  • Future work includes generalizing the scale-relationship model, developing more accurate statistical models of social-media noise, and improving algorithm scalability.Source: §9 · p. 30
Loading 1404.7048v2…