基于情感时间序列的微博热点主题检测

Tengjiao Wang, Binyang Li, Baolin Peng, Jing Li, Kam‐Fai Wong, Xu HanScientia 2015 · DOI

Abstract

Identifying trending topics on social media platforms like Twitter and Sina Weibo is crucial for real-time news and opinion tracking. Traditional volume-based detection methods often suffer from significant lag and context blindness. This paper presents a novel Sentiment-Topic Detection (STD) model that incorporates a non-parametric sentiment time series feature. By analyzing collective emotional shifts, the model achieves an impressive 82.3% recall rate and detects trends an average of 25 minutes faster than traditional methods.

The Problem: The Lag of Volume

In the high-velocity world of microblogging, "what's hot" changes in minutes. Traditional detection systems typically rely on "burst detection," which triggers alerts when keyword frequency or tweet volume spikes. However, these methods have two major flaws:

  1. Lag Time: By the time volume is high enough to be statistically significant, the event is often already old news.
  2. Context Blindness: High volume can be caused by spam or routine chatter rather than a genuine trending event.

The researchers argue that human emotion is a leading indicator. A sudden shift in collective sentiment—outrage, joy, or surprise—often precedes a massive spike in post volume.

Methodology: The STD Model

The proposed Sentiment-Topic Detection (STD) model is a non-parametric supervised approach that fuses three distinct features:

  • Tweet Volume ($V$): The standard count of posts.
  • User Volume ($U$): The number of unique participants, used to filter out bot-driven spam.
  • Sentiment Intensity ($S$): The core innovation of this research.

Creating the Sentiment Time Series

To transform messy, informal tweets into a usable signal, the authors employed a multi-stage pipeline. First, an SVM classifier trained on 1.6 million tweets scores individual posts on a polarity scale. Next, the researchers borrowed Histogram Equalization from image processing to enhance the contrast of these sentiment scores. Finally, Parzen Window Estimation with a Gaussian kernel was used to smooth the erratic raw data into a continuous curve, filtering out individual "jitter" to reveal the underlying emotional trend.

Visualizing the sentiment and volume time series signals used for detection.

Key Results and Performance

The model was validated using massive datasets: 900 million tweets from Twitter and three months of data from Sina Weibo. The results demonstrated that the integration of sentiment features significantly boosts detection capabilities.

Superior Recall and Speed

As shown in the performance data, the combined $V+S$ model achieved a recall rate of 82.3% on Twitter, outperforming models that rely solely on volume or user counts. Furthermore, the STD model identified trending topics an average of 25 minutes faster than baseline methods, occasionally beating Twitter’s own internal trending algorithm by approximately 5 minutes.

Comparative performance of different model configurations on Twitter and Sina Weibo datasets.

Cross-Platform Robustness

The model proved effective across both English (Twitter) and Chinese (Sina Weibo) platforms. This suggests that the link between emotional shifts and trending behavior is a universal human trait rather than a platform-specific quirk. Interestingly, the researchers found that while combining Volume ($V$) and Sentiment ($S$) yielded the best results, adding User Volume ($U$) sometimes introduced enough noise to slightly decrease performance, as $V$ and $U$ are often highly correlated.

Conclusion and Future Directions

The STD model proves that sentiment is a powerful, early-warning signal for social media trends. While the current system faces challenges with sarcasm and high computational overhead, it sets a new benchmark for real-time topic detection. Future work will likely explore deep learning architectures to better handle the nuances of informal language and improve the precision of the sentiment signal.

Summary updated

Abstract

Twitter plays a significant role in information diffusion, and it has evolved to become an important information resource as well as news feed. There is a widespread interest in what is happening on Twitter, and the instantaneous news information that is passed on. However, with the large amount of data, it is impossible to manually determine what topic is trending, which makes real-time topic detection attractive and significant. Furthermore, Twitter provides a platform for the sharing of opinions and providing feedback for events, news, and products, etc. Because users tend to express their real thoughts on Twitter, it is recognized as a valuable source of opinions. Nevertheless, most works about trending topic detection fail to consider sentiments. In this work, we develop a non-parametric supervised real-time-trending topic-detection model with a sentimental feature. By performing experiments, we show that our model successfully detects trending sentimental topic in a short time. After applying a combination of multiple features, e.g., tweet volume and user volume, the proposed model demonstrates impressive effectiveness with an 82.3% recall rate, surpassing all of the competitors.
BibTeX
@inproceedings{wang2015,
  author = {Tengjiao Wang and Binyang Li and Baolin Peng and Jing Li and Kam‐Fai Wong and Xu Han},
  title = {{基于情感时间序列的微博热点主题检测}},
  booktitle = {Scientia},
  year = {2015},
  doi = {10.1360/n112015-00249},
}
Figure