Data Mining and Knowledge Discovery

Chao Zhang, Jiawei Han2021 · DOI · OpenAlex

Abstract

Our physical world is being projected into online cyberspace at an unprecedented rate. People nowadays visit different places and leave behind them million-scale digital traces such as tweets, check-ins, Yelp reviews, and Uber trajectories. Such digital data are a result of social sensing: namely people act as human sensors that probe different places in the physical world and share their activities online. The availability of massive social-sensing data provides a unique opportunity for understanding urban space in a data-driven manner and improving many urban computing applications, ranging from urban planning and traffic scheduling to disaster control and trip planning. In this chapter, we present recent developments in data-mining techniques for urban activity modeling, a fundamental task for extracting useful urban knowledge from social-sensing data. We first describe traditional approaches to urban activity modeling, including pattern discovery methods and statistical models. Then, we present the latest developments in multimodal embedding techniques for this task, which learns vector representations for different modalities to model people's spatiotemporal activities. We study the empirical performance of these methods and demonstrate how data-mining techniques can be successfully applied to social-sensing data to extract actionable knowledge and facilitate downstream applications.

Introduction: The Rise of Social Sensing

Urban planning and traffic management have historically relied on expensive, slow, and coarse-grained surveys. However, the digital revolution has introduced "social sensing"—the million-scale digital traces left by citizens via tweets, check-ins, and trajectories. These traces act as human-generated probes into the physical state of a city.

Spatial distribution of social-sensing data points across an urban area.

While this data is abundant, it is notoriously "messy." It comprises disparate modalities: precise GPS coordinates, categorical timestamps, and unstructured text snippets. The central challenge in urban computing is multimodal correlation: building models that understand, for instance, that the keyword "flight" at 6:00 PM is intrinsically linked to the geographic coordinates of an airport. Traditional generative models, such as Latent Dirichlet Allocation (LDA), often struggle with the scale of this data or rely on rigid distributional assumptions that result in "blurry" predictions.

Methodology: The CrossMap Approach

To overcome the limitations of traditional statistical models, the authors propose CrossMap, a multimodal embedding technique. Unlike older "topic models," CrossMap maps location, time, and text into a shared high-dimensional latent space.

The CrossMap architecture: Integrating spatial, temporal, and textual data into a shared embedding space.

The CrossMap framework operates on four key principles:

  1. Shared Latent Space: Every urban region, hour of the day, and keyword is represented as a vector. Related attributes are pushed closer together in this mathematical space.
  2. Attribute Reconstruction: The model is trained via an unsupervised task: given two attributes (e.g., time and location), it must predict the third (e.g., the text).
  3. Spatiotemporal Smoothing: Addressing data sparsity, the model incorporates the "First Law of Geography"—the principle that near things are more related than distant ones. Spatial and temporal smoothing ensure that neighboring regions and consecutive hours maintain similar representations.
  4. Efficient Optimization: By utilizing Stochastic Gradient Descent (SGD) and negative sampling, CrossMap remains computationally efficient even when processing millions of data points.

Experimental Setup and Datasets

The efficacy of CrossMap was evaluated using three massive real-world datasets:

  • Los Angeles (Twitter): 1.10 million geo-tagged tweets (Aug–Nov 2014).
  • New York City (Twitter): 1.20 million geo-tagged tweets (Aug–Nov 2014).
  • Foursquare (4SQ): 0.7 million check-ins in NYC (2010–2011), featuring explicit venue categories like "Nightlife Spot" for activity classification.

Key Results and Insights

The study utilized Mean Reciprocal Rank (MRR) to evaluate the model's ability to rank correct attributes. The results demonstrated a paradigm shift in performance:

  • Predictive Accuracy: CrossMap achieved up to a 62% improvement in location prediction and an 83% improvement in activity prediction over traditional geographic topic models (LGTA and MGTM).
  • Classification Excellence: In downstream tasks, such as identifying activity types (e.g., "Travel & Transport"), CrossMap reached an F1 score of 0.843, significantly outperforming SVD and Tensor factorization methods.
  • Semantic Depth: Qualitative analysis showed that the model understands context. A query for "restaurant" at 10:00 AM retrieves breakfast-related keywords, while the same query at 8:00 PM yields dinner and "dating" terms.

Furthermore, t-SNE visualizations revealed that CrossMap creates much sharper clusters for distinct activities compared to the overlapping, messy results produced by older geographic topic models.

Conclusion and Future Directions

While CrossMap represents a significant leap forward, challenges remain. Approximately 40% of social media data is "pointless babble," making signal extraction difficult. Future research will focus on integrating visual modalities (images) to provide richer context and developing on-line learning algorithms to handle real-time urban events like disasters or protests. By treating citizens as sensors, we can transform digital noise into actionable urban intelligence.

Summary updated

Abstract

Abstract Our physical world is being projected into online cyberspace at an unprecedented rate. People nowadays visit different places and leave behind them million-scale digital traces such as tweets, check-ins, Yelp reviews, and Uber trajectories. Such digital data are a result of social sensing: namely people act as human sensors that probe different places in the physical world and share their activities online. The availability of massive social-sensing data provides a unique opportunity for understanding urban space in a data-driven manner and improving many urban computing applications, ranging from urban planning and traffic scheduling to disaster control and trip planning. In this chapter, we present recent developments in data-mining techniques for urban activity modeling, a fundamental task for extracting useful urban knowledge from social-sensing data. We first describe traditional approaches to urban activity modeling, including pattern discovery methods and statistical models. Then, we present the latest developments in multimodal embedding techniques for this task, which learns vector representations for different modalities to model people's spatiotemporal activities. We study the empirical performance of these methods and demonstrate how data-mining techniques can be successfully applied to social-sensing data to extract actionable knowledge and facilitate downstream applications.
BibTeX
@inproceedings{zhang2021data,
  author = {Chao Zhang and Jiawei Han},
  title = {{Data Mining and Knowledge Discovery}},
  booktitle = {˜The},
  year = {2021},
  doi = {10.1007/978-981-15-8983-6\_42},
}
Figure