Christian Pölitz

Discovering bits of place histories from people's activity traces

Events that happened in the past are important for understanding the ongoing processes, predicting future developments, and making informed decisions. Significant and/or interesting events tend to attract many people. Some people leave traces of their attendance in the form of computer-processable data, such as records in the databases of mobile phone operators or photos on photo sharing web sites. We developed a suite of visual analytics methods for reconstructing past events from these activity traces. Our tools combine geocomputations, interactive geovisualizations and statistical methods to enable integrated analysis of the spatial, temporal, and thematic components of the data, including numeric attributes and texts. We demonstrate the utility of our approach on two large real data sets, mobile phone calls in Milano during 9 days and flickr photos made on British Isles during 5 years.

A framework for using self-organising maps to analyse spatio-temporal patterns, exemplified by analysis of mobile phone usage

We suggest a visual analytics framework for the exploration and analysis of spatially and temporally referenced values of numeric attributes. The framework supports two complementary perspectives on spatio-temporal data: as a temporal sequence of spatial distributions of attribute values (called spatial situations) and as a set of spatially referenced time series of attribute values representing local temporal variations. To handle a large amount of data, we use the self-organising map (SOM) method, which groups objects and arranges them according to similarity of relevant data features. We apply the SOM approach to spatial situations and to local temporal variations and obtain two types of SOM outcomes, called space-in-time SOM and time-in-space SOM, respectively. The examination and interpretation of both types of SOM outcomes are supported by appropriate visualisation and interaction techniques. This article describes the use of the framework by an example scenario of data analysis. We also discuss how the framework can be extended from supporting explorative analysis to building predictive models of the spatio-temporal variation of attribute values. We apply our approach to phone call data showing its usefulness in real-world analytic scenarios.

Identifying Place Histories from Activity Traces with an Eye to Parameter Impact

Events that happened in the past are important for understanding the ongoing processes, predicting future developments, and making informed decisions. Important and/or interesting events tend to attract many people. Some people leave traces of their attendance in the form of computer-processable data, such as records in the databases of mobile phone operators or photos on photo sharing web sites. We developed a suite of visual analytics methods for reconstructing past events from these activity traces. Our tools combine geocomputations, interactive geovisualizations, and statistical methods to enable integrated analysis of the spatial, temporal, and thematic components of the data, including numeric attributes and texts. We also support interactive investigation of the sensitivity of the analysis results to the parameters used in the computations. For this purpose, statistical summaries of computation results obtained with different combinations of parameter values are visualized in a way facilitating comparisons. We demonstrate the utility of our approach on two large real data sets, mobile phone calls in Milano during 9 days and flickr photos made on British Isles during 5 years.

Adaptive burst detection in a stream engine

Detecting bursts in data streams is an important and challenging task. Due to the complexity of this task, usually burst detection cannot be formulated using standard query operators. Therefore, we show how to integrate burst detection for stationary as well as non-stationary data into query formulation and processing, from the language level to the operator level. Afterwards, we present fundamentals of threshold-based burst detection. We focus on the applicability of time series forecasting techniques in order to dynamically identify suitable thresholds for stream data containing arbitrary trends and periods. The proposed approach is evaluated with respect to quality and performance on synthetic and real-world sensor data using a full-fledged DSMS.

Finding Arbitrary Shaped Clusters with Related Extents in Space and Time

The paper deals with density-based clustering of events, i.e. objects positioned in space and time, such as occurrences of earthquakes, forest fires, mobile phone calls, or photos taken by Flickr users. Finding concentrations of events in space and time can help to discover interesting places and time periods. The spatial and temporal properties of event clusters, in particular, their spatial and temporal extents and densities, can be related to each other. According to Tobler’s Law, the relationship can be described as follows. Events in a small area can be sparse in time and still connected. On the other hand, events in large areas are likely to be connected if they are close in time. Hence, the temporal distance threshold for density-based clustering should vary depending on the spatial extent of the area in which events happen. Therefore, we suggest a two-step clustering method. In the first step, spatial clusters of events are detected. In the second step, density-based clustering is applied to the temporal positions of spatially clustered events. The temporal distance threshold is chosen individually for each spatial cluster depending on its spatial extent. We demonstrate the work of the method on several examples of real data.

Interpretable domain adaptation via optimization over the Stiefel manifold

In domain adaptation, the goal is to find common ground between two, potentially differently distributed, data sets. By finding common concepts present in two sets of words pertaining to different domains, one could leverage the performance of a classifier for one domain for use on the other domain. We propose a solution to the domain adaptation task, by efficiently solving an optimization problem through Stochastic Gradient Descent. We provide update rules that allow us to run Stochastic Gradient Descent directly on a matrix manifold: the steps compel the solution to stay on the Stiefel manifold. This manifold encompasses projection matrices of word vectors onto low-dimensional latent feature representations, which allows us to interpret the results: the rotation magnitude of the word vector projection for a given word corresponds to the importance of that word towards making the adaptation. Beyond this interpretability benefit, experiments show that the Stiefel manifold method performs better than state-of-the-art methods.

Learning to rank under tight budget constraints

This paper investigates the influence of pruning feature lists to keep a given budget for the evaluation of ranking methods. We learn from a given training set how important the individual prefixes are for the ranking quality. Based on there importance we choose the best prefixes to calculate the ranking while keeping the budget.

Investigation of Word Senses over Time Using Linguistic Corpora

Word sense induction is an important method to identify possible meanings of words. Word co-occurrences can group word contexts into semantically related topics. Besides the pure words, temporal information provide another dimension to further investigate the development of the word meanings over time. Large digital corpora of written language, such as those that are held by the CLARIN-D centers, provide excellent possibilities for such kind of linguistic research on authentic language data. In this paper, we investigate the evolution of meanings of words with topic models over time using large digital text corpora.

Supervised Extraction of Usage Patterns in Different Document Representations

Finding usage patterns of words in documents is an important task in language processing. Latent topic or latent factor models can be used to find hidden connections between words in documents based on correlations among them. Depending on the representation of the documents, correlations between different elements can be found. Given additional labels (either numeric or nominal) for the documents, we can further infer the usage patterns that reflect this given information. We present an empirical comparison of topic and factor models for different documents representations to find usage patterns of words in a large document collections that explain given label information.

Distance Based Active Learning for Domain Adaptation

We investigate methods to apply Domain Adaptation coupled with Active Learning to reduce the number of labels needed to train a classifier. We assume to have a classification task on a given unlabelled set of documents and access to labels from different documents of other sets. The documents from the other sets come from different distributions. Our approach uses Domain Adaptation together with Active Learning to find a minimum number of labelled documents from the different sets to train a high quality classifier. We assume that documents from different sets that are close in a latent topic space can be used for a classification task on a given different set of documents.