Skip to main content

Fries Jason

Emerging event detection is the process of automatically identifying novel and emerging ideas from text with minimal human intervention. With the rise of social networks like Twitter, topic detection has begun leveraging measures of user influence to identify emerging events. Twitter's highly skewed follower/followee structure lends itself to an intuitive model of influence, yet in a context like the Emerging Infections Network (EIN), a sentinel surveillance listserv of over 1400 infectious disease experts, developing a useful model of authority becomes less clear. Who should we listen to on the EIN? To explore this, we annotated a body of important EIN discussions and tested how well 3 models of user authority performed in identifying those discussions. In previous work we proposed a process by which only posts that are based on specific "important" topics are read, thus drastically reducing the amount of posts that need to be read. The process works by finding a set of "bellwether" users that act as indicators for "important" topics and only posts relating to these topics are then read. This approach does not consider the text of messages, only the patterns of user participation. Our text analysis approach follows that of Cataldi et al.[1], using the idea of semantic "energy" to identify emerging topics within Twitter posts. Authority is calculated via PageRank and used to weight each author's contribution to the semantic energy of all terms occurring in within some interval ti. A decay parameter d defines the impact of prior time steps on the current interval.

Objective

To explore how different models of user influence or authority perform when detecting emerging events within a small-scale community of infectious disease experts.

Referenced File
Submitted by elamb on

The increasing use of the Internet to arrange sexual encounters presents challenges to public health agencies formulating STD interventions, particularly in the context of anonymous encounters. These encounters complicate or break traditional interventions. In previous work [1], we examined a corpus of anonymous personal ads seeking sexual encounters from the classifieds website Craigslist and presented a way of linking multiple ads posted across time to a single author. The key observation of our approach is that some ads are simply reposts of older ads, often updated with only minor textual changes. Under the presumption that these ads, when not spam, originate from the same author, we can use efficient near-duplicate detection techniques to cluster ads within some threshold similarity. Linking ads in this way allows us to preserve the anonymity of authors while still extracting useful information on the frequency with which authors post ads, as well as the geographic regions in which they seek encounters. While this process detects many clusters, the lack of a true corpus of authorship-linked ads makes it difficult to validate and tune the parameters of our system. Fortunately, many ad authors provide an obfuscated telephone number in ad text (e.g., 867-5309 becomes 8sixseven5three oh nine) to bypass Craigslist filters, which prohibit including phone numbers in personal ads. By matching phone numbers of this type across all ads, we can create a corpus of ad clusters known to be written by a single author. This authorship corpus can then be used to evaluate and tune our existing near-duplicate detection system, and in the future identify features for more robust authorship attribution techniques.

Objective:

This paper constructs an authorship-linked collection or corpus of anonymous, sex-seeking ads found on the classifieds website Craigslist. This corpus is then used to validate an authorship attribution approach based on identifying near duplicate text in ad clusters, providing insight into how often anonymous individuals post sexseeking ads and where they meet for encounters.

Submitted by Magou on
Subscribe to Fries Jason