AN APPROACH TO USER FEEDBACK PROCESSING ORDER TO INCREASE QUALITY OF CLUSTERING RESULTS

Dataset clustering could have more than one “right” result depending on a user intention. For example, texts could be clustered according to their topic, style or author. In case of unsatisfactory results, a data scientist needs to re-construct a feature space in order to change the results. The relation between the feature space and the result are often quite complicated. The latter results in building several clustering models to explore useful relations. Interactive clustering with feedback is aimed to cope with this problem. In this paper an approach to user feedback processing during clustering is presented. The approach is based on end-to-end clustering and uses an autoencoder neural network. This technique allows to adjust iteratively the computing clusters without changing feature space.

Interactive trace clustering

Process mining aims to automatically discover, analyze and improve business processes. Trace clustering is a task commonly used to reduce the inherent complexity of processes by identifying patterns. This research focuses on the application of experts knowledge in process mining through interactive clustering, referred to herein as interactive trace clustering. The aim is to improve trace clustering by reducing potential losses arising from arbitrary assumptions on the similarity between the datapoints, what is commonly required in unsupervised scenarios. Initial experiments considered partitioning clustering and three representation schemes for traces. Preliminary results show potential to improve the trace clustering quality by inserting experts knowledge.