Christian Sengstock

EvenTweet: online localized event detection from twitter

Microblogging services such as Twitter, Facebook, and Foursquare have become major sources for information about real-world events. Most approaches that aim at extracting event information from such sources typically use the temporal context of messages. However, exploiting the location information of georeferenced messages, too, is important to detect localized events, such as public events or emergency situations. Users posting messages that are close to the location of an event serve as human sensors to describe an event. In this demonstration, we present a novel framework to detect localized events in real-time from a Twitter stream and to track the evolution of such events over time. For this, spatio-temporal characteristics of keywords are continuously extracted to identify meaningful candidates for event descriptions. Then, localized event information is extracted by clustering keywords according to their spatial similarity. To determine the most important events in a (recent) time frame, we introduce a scoring scheme for events. We demonstrate the functionality of our system, called Even-Tweet, using a stream of tweets from Europe during the 2012 UEFA European Football Championship.

Latent geographic feature extraction from social media

In this work we present a framework for the unsupervised extraction of latent geographic features from georeferenced social media. A geographic feature represents a semantic dimension of a location and can be seen as a sensor that measures a signal of geographic semantics. Our goal is to extract a small number of informative geographic features from social media, to describe and explore geographic space, and for subsequent spatial analysis, e.g., in market research. We propose a framework that, first, transforms the unstructured and noisy geographic information in social media into a high-dimensional multivariate signal of geographic semantics. Then, we use dimensionality reduction to extract latent geographic features. We conduct experiments using two large-scale Flickr data sets covering the LA area and the US. We show that dimensionality reduction techniques extracting sparse latent features find dimensions with higher informational value. In addition, we show that prior normalization can be used as a parameter in the exploration process to extract features representing different geographic characteristics, that is, landmarks, regional phenomena, or global phenomena.

CONQUER: a system for efficient context-aware query suggestions

Many of todays search engines provide autocompletion while the user is typing a query string. This type of dynamic query suggestion can help users to formulate queries that better represent their search intent during Web search interactions. In this paper, we demonstrate our query suggestion system called CONQUER, which allows to efficiently suggest queries for a given partial query and a number of available query context observations. The context-awareness allows for suggesting queries tailored to a given context, e.g., the user location or the time of day. CONQUER uses a suggestion model that is based on the combined probabilities of sequential query patterns and context observations. For this, the weight of a context in a query suggestion can be adjusted online, for example, based on the learned user behavior or user profiles. We demonstrate the functionality of CONQUER based on 6 million queries from an AOL query log using the time of day and the country domain of the clicked URLs in the search result as context observations.

Spatio-temporal characteristics of bursty words in Twitter streams

Social networking and microblogging services such as Twitter provide a continuous source of data from which useful information can be extracted. The detection and characterization of bursty words play an important role in processing such data, as bursty words might hint to events or trending topics of social importance upon which actions can be triggered. While there are several approaches to extract bursty words from the content of messages, there is only little work that deals with the dynamics of continuous streams of messages, in particular messages that are geo-tagged. In this paper, we present a framework to identify bursty words from Twitter text streams and to describe such words in terms of their spatio-temporal characteristics. Using a time-aware word usage baseline, a sliding window approach over incoming tweets is proposed to identify words that satisfy some burstiness threshold. For these words then a time-varying, spatial signature is determined, which primarily relies on geo-tagged tweets. In order to deal with the noise and the sparsity of geo-tagged tweets, we propose a novel graph-based regularization procedure that uses spatial cooccurrences of bursty words and allows for computing sound spatial signatures. We evaluate the functionality of our online processing framework using two real-world Twitter datasets. The results show that our framework can efficiently and reliably extract bursty words and describe their spatio-temporal evolution over time.

Spatial Interestingness Measures for Co-location Pattern Mining

Co-location pattern mining aims at finding subsets of spatial features frequently located together in spatial proximity. The underlying motivation is to model the spatial correlation structure between the features. This allows to discover interesting co-location rules (feature interactions) for spatial analysis and prediction tasks. As in association rule mining, a major problem is the huge amount of possible patterns and rules. Hence, measures are needed to identify interesting patterns and rules. Existing approaches so far focused on finding frequent patterns, patterns including rare features, and patterns occurring in small (local) regions. In this paper, we present a new general class of interestingness measures that are based on the spatial distribution of co-location patterns. These measures allow to judge the interestingness of a pattern based on properties of the underlying spatial feature distribution. The results are different from standard measures like participation index or confidence. To demonstrate the usefulness of these measures, we apply our approach to the discovery of rules on a subset of the OpenStreetMap point-of-interest data.

A probablistic model for spatio-temporal signal extraction from social media

It is nowadays possible to access a huge and increasing stream of social media records. Recently, such data has been used to infer about spatio-temporal phenomena by treating the records as proxy observations of the real world. However, since such observations are heavily uncertain and their spatio-temporal distribution is highly heterogeneous, extracting meaningful signals from such data is a challenging task. In this paper, we present a probabilistic model to extract spatio-temporal distributions of phenomena (called spatio-temporal signals) from social media. Our approach models spatio-temporal and semantic knowledge about real-world phenomena embedded in records on the basis of conditional probability distributions in a Bayesian network. Through this, we realize a generic and comprehensive model where knowledge and uncertainties about spatio-temporal phenomena can be described in a modular and extensible fashion. We show that existing models for the extraction of spatio-temporal phenomena distributions from social media are particular instances of our model. We quantitatively evaluate instances of our model by comparing the spatio-temporal distributions of extracted phenomena from a large Twitter data set to their real-world distributions. The results clearly show that our model allows to extract better spatio-temporal signals in terms of quality and robustness.

Exploration and comparison of geographic information sources using distance statistics

Given the steadily increasing amount of geographic information on the Web, there is a strong need for suitable methods in exploratory data analysis that can be used to efficiently describe the characteristics of such large-scale, often noisy datasets. Existing methods in spatial data mining focus primarily on mining patterns describing spatial proximity relationships such as co-location patterns or spatial associations rules. In this paper, we present a novel approach to describe the spatial characteristics of geographic information sources comprised of instances of geographic features. Using the concept of interaction characteristics of geographic features, similarities in how features are distributed in space can be computed and interesting patterns of similar features in the datasets regarding their geographic semantics (landmark, local, regional, global) can be determined. For this, we employ clustering techniques of spatial distance statistics. We discuss the properties of our method and detail a comprehensive evaluation using publicly available datasets (Flickr, Twitter, OpenStreeMap). We demonstrate the feasibility of identifying groups of geographic features with distinct geographic semantics, which then can be used to select subsets of features for subsequent learning tasks or to compare different datasets.

Reliable spatio-temporal signal extraction and exploration from human activity records

Shared multimedia, microblogs, search engine queries, user comments, and location check-ins, among others, generate an enormous stream of human activity records. Such records consist of information in the form of text, images, or videos, and can often be traced in time and space using associated time/location information. Over the past years such spatio-temporal activity streams have been heavily studied with the aim to extract and explore spatio-temporal phenomena, like events, place descriptions, and geographical topics. Despite the clear intuition and often simple techniques to extract such knowledge, the amount of noise, sparsity, and heterogeneity in the data makes such tasks non-trivial and erroneous. This demonstration offers a visual interface to compare, combine, and evaluate spatio-temporal signal extraction and exploration approaches from large-scale sets of human activity records.

An Event-Based Framework for the Semantic Annotation of Locations

There is an increasing number of Linked Open Data sources that provide information about geographic locations, e.g., GeoNames or LinkedGeoData. There are also numerous data sources managing information about events, such as concerts or festivals. Suitably combining such sources would allow to answer queries such as ‘When and where do live-concerts most likely occur in Munich?’ or ‘Are two locations similar in terms of their events?’. Deriving correlations between geographic locations and event data, at different levels of abstraction, provides a semantically rich basis for location search, topic-based location clustering or recommendation services. However, little work has been done yet to extract such correlations from event datasets to annotate locations.

Spatial Itemset Mining: A Framework to Explore Itemsets in Geographic Space

Driven by the major adoption of mobile devices, user contributed geographic information has become ubiquitous. A typical example is georeferenced and tagged social media, linking a location to a set of features or attributes. Mining frequent sets of discrete attributes to discover interesting patterns and rules of attribute usage in such data sets is an important data mining task. In this work we extend the frequent itemset mining framework to model the spatial distribution of itemsets and association rules. For this, we expect the input transactions to have an associated spatial attribute, as, for example, present in georeferenced tag sets. Using the framework, we formulate interestingness measures that are based on the underlying spatial distribution of the input transactions, namely area, spatial support, location-conditional support, and spatial confidence. We show that describing the spatial characteristics of itemsets cannot be handled by existing approaches to mine association rules with numeric attributes, and that the problem is different from co-location pattern mining and spatial association rules mining. We demonstrate the usefulness of our proposed extension by different mining tasks using a real-world data set from Flickr.