Inhaltszusammenfassung

Traditional pattern mining algorithms require access to the data, either in the form of a complete set of data, as in batch data mining, or in the form of a window of recent data, as in stream mining. In the case of stream mining, this comes with a number of disadvantages, such as the possibly unbounded growth of relevant instances, drift, possibly changing data mining tasks, and issues with privacy, to name a few. Therefore, an approach has been recently proposed that extracts patterns just ...Traditional pattern mining algorithms require access to the data, either in the form of a complete set of data, as in batch data mining, or in the form of a window of recent data, as in stream mining. In the case of stream mining, this comes with a number of disadvantages, such as the possibly unbounded growth of relevant instances, drift, possibly changing data mining tasks, and issues with privacy, to name a few. Therefore, an approach has been recently proposed that extracts patterns just from statistical information of the stream - more precisely, an online density estimate that is inferred from it. As this approach is mainly based on sampling from the density estimates, it still struggles with itemsets having a medium to low frequency. To resolve this issue, we pursue an alternative strategy in this paper and directly exploit the structure of the density estimates to extract frequent itemsets. Additionally, we address the important matter of privacy-preserving data mining by ensuring that the density estimate fulfills privacy-related properties. To show the effectiveness of the proposed methods, we provide proofs and evaluate the performance on synthetic and real-world data.» weiterlesen» einklappen

Autoren

Klassifikation

DFG Fachgebiet:
Informatik

DDC Sachgruppe:
Informatik