Abstract | ||
---|---|---|
This paper presents a novel approach for the detection of contextual anomalies. This approach, based on log-linear tensor factorization, considers a stream of discrete events, each representing the co-occurence of contextual elements, and detects events with low-probability. A parametric model is used to learn the joint probability of contextual elements, in which the parameters are the factors of the event tensor. An efficient method, based on Nesterov's accelerated gradient ascent, is proposed to learn these parameters. The proposed approach is evaluated on the low-rank approximation of tensors, the prediction of future of events and the detection of events representing abnormal behaviors. Results show our method to outperform state of the art approaches for these problems. |
Year | DOI | Venue |
---|---|---|
2015 | 10.1007/978-3-319-18032-8_13 | ADVANCES IN KNOWLEDGE DISCOVERY AND DATA MINING, PART II |
Keywords | Field | DocType |
Contextual anomaly detection,Tensor factorization,Low-rank approximation,Future event prediction | Anomaly detection,Data mining,Gradient descent,Parametric model,Joint probability distribution,Tensor,Computer science,Low-rank approximation,Artificial intelligence,Log-linear model,Tensor factorization,Machine learning | Conference |
Volume | ISSN | Citations |
9078 | 0302-9743 | 0 |
PageRank | References | Authors |
0.34 | 18 | 3 |
Name | Order | Citations | PageRank |
---|---|---|---|
Alpa Jayesh Shah | 1 | 0 | 0.34 |
Christian Desrosiers | 2 | 100 | 23.90 |
Robert Sabourin | 3 | 908 | 61.89 |