Multidimensional Anomaly Detection

[seanlaw]

There is a new paper here

[SaVoAMP]

Hi Sean,

I'm not sure if you still remember me, but back when I was working on my bachelor's thesis, I spent quite some time studying matrix profiles for multidimensional motif discovery. We had several discussions about the approach, and I also tried to contribute a bit to the implementation.

In the meantime, I have completed my master's thesis, though I still need to finish one last project before officially graduating. I started this project alongside my work as a student assistant at the Fraunhofer Institute. The goal is to compare different anomaly detection methods on time series data. For this project, I have been diving deep into machine learning methods, and the process has been dragging on because I've frequently encountered questionable results and inconsistencies in many ML approaches. Interestingly, a bachelor’s student at our university is facing similar issues, and our professor has even suggested writing a paper about the problems and questionable practices we've come across.

Long story short: while writing to my professor about this, I came across the comparatively new paper from the matrix profile group on multidimensional time series analysis, and it piqued my curiosity. I immediately checked the STUMPY repository to see if anything has been implemented yet and found this issue.

Have you already had a chance to take a closer look at the paper or perhaps implemented anything from it? I haven't read the paper in detail yet, but I'm quite interested, mainly because I always found the matrix profile papers to be of much higher quality than most of the work I've encountered in the machine learning field so far.

Also, from what I've seen, most ML-based anomaly detection methods tend to identify points in time where anomalies occur, but they rarely provide concrete information about in which dimensions or how many dimensions the anomalies manifest. This leads exactly to the kinds of problems the matrix profile group described: when searching across all dimensions simultaneously, there is a risk of missing individual motifs (or anomalies), while treating each dimension separately can ignore important interdependencies between dimensions that are critical for identifying anomalies.

Best regards,
Sarah

[seanlaw]

I'm not sure if you still remember me

Yes, of course I remember you @SaVoAMP! I hope all is well and congratulations on your Masters thesis.

Have you already had a chance to take a closer look at the paper or perhaps implemented anything from it?

No, we haven't had a chance to study this yet as other tasks have taken a higher priority. Having said that, we are certainly interested in at least confirming, as a first step, that the paper is reproducible. Will you consider creating a notebook that could possibly reproduce the work?

I always found the matrix profile papers to be of much higher quality than most of the work I've encountered in the machine learning field so far.

I completely agree!

[SaVoAMP]

Thank you!

Will you consider creating a notebook that could possibly reproduce the work?

I'm definitely interested, but I first need to focus on completing my current project to finish my degree. Time is starting to run a bit short, especially since I'm also continuing to contribute to an open-source framework I was involved with during my master's thesis, as we are aiming to publish a paper based on that work.

From the end of July or early August, I should have a bit more time available. I could perhaps read through the paper in the coming days and get an initial overview.

If you like, we could proceed as follows: you could tentatively keep me in mind, but I can't guarantee how quickly I’ll be able to make progress or how much I’ll be able to contribute in the near term. I would work on it whenever time allows. However, if someone else is motivated and would like to take over the issue in the meantime, they are of course welcome to do so.

What do you think?

[seanlaw]

That sounds like a good plan! Please don't feel pressured to take this on if you already have a lot on your plate. There is absolutely no rush to get started but I'd like to resolve it promptly once we choose to commence (and not let it drag out in time). Thanks again for thinking of us!

[SaVoAMP]

That sounds very reasonable. I'll take a look at the paper when I find a bit of time and will let you know once I have a better idea of whether I can contribute.