Links und Funktionen
Sprachumschaltung

Navigationspfad
Sie sind hier: Startseite / Lehre / SS 2019 / Oberseminar / Oberseminar TCS Planungssitzung und Ankuendigung -- Julian Jorczik: Natural language processing and unsupervised machine learning for automated log analysis of mobile radio test systems


Inhaltsbereich

Oberseminar TCS Planungssitzung und Ankuendigung -- Julian Jorczik: Natural language processing and unsupervised machine learning for automated log analysis of mobile radio test systems

Julian Jorczik: Natural language processing and unsupervised machine learning for automated log analysis of mobile radio test systems
Wann 14:15 16:00 26.04.2019
von bis
Wo Raum L109, Oettingenstr. 67
Termin übernehmen vCal
iCal

Es spricht Julian Jorczik über:
Natural language processing and unsupervised machine learning for automated log analysis of mobile radio test systems

 

Abstract:
Using machine learning to analyze logs represents a promising way to acquire
valuable information about complex processes on distributed systems without
the need for human intervention. Current research shows that it is not clear
yet how to combine log parsing and log mining for effective log analysis.
This talk examines how to apply machine learning to analyze logs of mobile
radio test systems when labeled data is not available. To pursue this
question, the study specifies applications of analyzing logs of mobile radio
test systems. The study also reviews the recent progress that has been made
in log analysis in the specific area of research. Beyond that, the study
elaborates prototypes based on natural language processing and unsupervised
machine learning for error diagnosis and anomaly detection. The prototype
for error diagnosis represents a text clustering approach that achieves
competitive results in terms of purity (0.953), adjusted rand index (0.890)
and normalized mutual information (0.861) on a real world benchmark dataset
for document clustering in comparison with state-of-the-art methods. The
evaluation shows that the approach effectively diagnoses known log message
sequences, but struggles for diagnosing unknown log message sequences. The
prototype for anomaly detection exploits a density-based outlier detection
algorithm for computing probabilities of logs being anomalous. The approach
achieves excellent results on a test dataset in terms of precision, with
values above 90% for specific settings. This presentation shows that it is
possible to analyze logs effectively by combining techniques of natural
language processing and unsupervised machine learning. Further
Investigation is required to identify improvements in both approaches.

Artikelaktionen


Funktionsleiste