Subspace Search in Data Streams

Aus SDQ-Institutsseminar
Version vom 12. November 2019, 16:05 Uhr von Edouard Fouché (Diskussion | Beiträge)
(Unterschied) ← Nächstältere Version | Aktuelle Version (Unterschied) | Nächstjüngere Version → (Unterschied)
Vortragende(r) Florian Kalinke
Vortragstyp Masterarbeit
Betreuer(in) Edouard Fouché
Termin Fr 22. November 2019
Kurzfassung Modern data mining often takes place on high-dimensional data streams, which evolve at a very fast pace: On the one hand, the "curse of dimensionality" leads to a sparsely populated feature space, for which classical statistical methods perform poorly. Patterns, such as clusters or outliers, often hide in a few low-dimensional subspaces. On the other hand, data streams are non-stationary and virtually unbounded. Hence, algorithms operating on data streams must work incrementally and take concept drift into account.

While "high-dimensionality" and the "streaming setting" provide two unique sets of challenges, we observe that the existing mining algorithms only address them separately. Thus, our plan is to propose a novel algorithm, which keeps track of the subspaces of interest in high-dimensional data streams over time. We quantify the relevance of subspaces via a so-called "contrast" measure, which we are able to maintain incrementally in an efficient way. Furthermore, we propose a set of heuristics to adapt the search for the relevant subspaces as the data and the underlying distribution evolves.

We show that our approach is beneficial as a feature selection method and as such can be applied to extend a range of knowledge discovery tasks, e.g., "outlier detection", in high-dimensional data-streams.