Student/in Informatik (w/m/d) - Masterthesis Spatial temporal bias challenge

Epidemiology extensively employs spatial and temporal models for predicting outbreaks and hotspots based on real historical data. In scenarios involving infrequent events like hotspots, a common challenge arises from the uneven distribution of classes in predictor variables. Standard solutions include techniques like stratified k-fold and oversampling or under sampling etc. However, when dealing with geospatial and temporal data, these techniques may introduce biases. This bias may emerge from oversampling data from the same geographical regions or similar timeframes. This, if not dealt with in care, can lead to a bias model.

Spatial models represent another problematic area since these models, incorporating methods like spatial autocorrelations etc., aim to address unbiased class distribution often overlook temporal factors. Conversely, temporal models focus on temporal aspects but may overlook spatial considerations.

In cases such as outbreaks and hotspots, it is crucial to incorporate biased data that reflects inherent spatial and temporal patterns.

A possible strategy to solve this problem is to preprocess data and fed into the model with some loss or cost function. This function can be calculated based on the gathered hotspot information from the neighboring data in the radius of each geographical point. Closer points show similar patterns can could hold the same hotspot value, farther points are more likely to be dissimilar. The points in between can get a continuous value based on its position within the radius. For this it is important to divide the space employing a hierarchical spatial representation, like a quadtree or kD-tree. Another important aspect is than based on each point further parameters can be imputed for example mobility information.

In this thesis, the student will have to undertake thorough gap analyses, delving into existing literature, experimenting with prevailing solutions, pinpointing disparities, and formulating approaches to address identified problems.

Students interested in this position should have a Computer Science background, proficient in Python and should have previous data analysis experience. An asset would be basic knowledge in the development of algorithms using spatial data structures, but this is not a must.