• Alexander de Linde Agerskov
  • Christian Bro Sørensen
  • Trine Juhl Holmager
4. term, Software, Master (Master Programme)
Knowing bacterial interactions in waste water treatment plants (WWTPs) can help manage the cleaning process, optimize it, avoid eutrophication and avoid pollution of effluent waters.
In this paper we propose to find bacterial interactions in WWTPs by clustering pairwise univariate time series consisting of bacterial abundances sampled from activated sludge.
We do this by modifying a deep clustering method called DPSOM to take pairs of bacteria as input.
We then propose to split these pairs into subsequences called windows and performing the clustering on these windows.
These cluster of windows are then used to produce clusters of the original full-length pairs.
To help understand the clustering of the pairs we provide visual explanations, with the LIME framework, of which features in a pair contribute to that pair's clustering.
As the dataset contains no ground truth in terms of interactions, we propose to evaluate our model using non-standard clustering metrics such as Pearson correlation coefficient and cluster-based prediction in addition to the LIME explanations.
Publication date11 Jun 2021
Number of pages18
ID: 414454263