Home

Igor Halperin

Likes
196
Date
2022/07/11
Image Link
https://media-exp2.licdn.com/dms/image/C5622AQG27zvN_t4Sqw/feedshare-shrink_800/0/1657500304502?e=1660176000&v=beta&t=GWooebJ5mGvBAGn3LzA6IvBfQoDU40EQ5sM0K1xxHkg
ML Score
4
Job
Fidelity Investments | AI Asset Management
Content
Last year my student and I wrote a paper pointing out the most time series anomaly detection benchmark datasets are worthless [a]. Among many other flaws, they are trivial. To make “trivial”, more formal, we define trivial as a problem that can be solved with one line of code (a poor man’s Kolmogorov Complexity!). What is very strange is that many papers now cite our paper but introduce new datasets that are just as flawed! The VLDB conference this year has at least two such papers [c].The figure shows an example, this multidimensional problem can be perfectly solved with one line of code: “ T(:,1)-T(:,2)>1 ”.Why are we doing this? It is possible to find real challenging datasets in this domain (it is a little more work, but not much) [b].A year after [a], the situation has gotten worse, not better. [a] Wu & Keogh: Current Time Series Anomaly Detection Benchmarks are Flawed and are Creating the Illusion of Progress.  arxiv.org/abs/2009.13807[b] https://lnkd.in/djeVJrFi[c] Anomaly Detection in Time Series: A Comprehensive Evaluation VLDB 2022 #VLDB #timeseries #anomalydetection
Property
Integromat
Link
https://www.linkedin.com/feed/update/urn:li:activity:6952182522201985024
Comments
17
Type
this
Profile
https://www.linkedin.com/in/igor-halperin-092175a/
TOP