Crime predictors have been sought after by governments and citizens alike for preventing or avoiding crimes. In this paper, we attempt to thoroughly analyze crime predictors from three Web open data sources: Google Street View (GSV), Twitter, and Foursquare, which provides visual, textual, and human behavioral data respectively. In contrast to existing works that attempt crime prediction at zip-code level or coarser granularity, we focus on street-level crime prediction. We transform data assigned to street-segments, and extract and determine strong predictors correlated with crime. Particularly, we are the first to discover visual clues on street outlooks that are predictive for crime. We focus on the city of San Francisco, and our extensive experiments show the effectiveness of predictors in a range of tests. We show that by analyzing and selecting strong predictors in Web open data, one could achieve significantly better crime prediction accuracy, comparing to traditional demographic data-based prediction.

中文翻译：

---

网络开放数据中街头犯罪预测因子分析

政府和公民都在寻求犯罪预测器来预防或避免犯罪。在本文中，我们试图彻底分析来自三个 Web 开放数据源的犯罪预测因素：Google 街景 (GSV)、Twitter 和 Foursquare，它们分别提供视觉、文本和人类行为数据。与尝试在邮政编码级别或更粗粒度的犯罪预测的现有工作相比，我们专注于街道级别的犯罪预测。我们转换分配给街道段的数据，并提取和确定与犯罪相关的强预测因子。特别是，我们是第一个发现可预测犯罪的街道景观视觉线索的人。我们专注于旧金山市，我们的大量实验在一系列测试中显示了预测器的有效性。