Create module for motif discovery

[mihailescum]

What do you think of having a separate module called motifs.py where we have an implementation of the k-top-motif search?

Do you think you could provide a checklist on the steps that you think are necessary to get this to work?

I would suggest the following:

• Leave core.mass as it is, because 8t does exactly what MASS was made for, i.e. finding the distance profile of a query.
• Rename stump.mass like you proposed
• In motifs.py have a function find_occurences (or something different) that does the following:

1. Gets a query Q, time series T and optionally the index of Q in T and an exclusion zone.
2. Replaces nan/inf in a copy of T and computes mean and stddev (basically does all the usual preprocessing)
3. Computes the distance profile of Q and T (the modified version) using core.mass, and if provided applies the exclusion zone.
4. Returns all indices below a threshold

• Then around this function (and by doing some caching, as to not do the preprocessing of T every time) one cans build a k motif search implementation.

Originally posted by @mexxexx in #166 (comment)

[seanlaw]

@mexxexx Did you happen to see this recent paper by Keogh's group on motif/discord discovery? It looks like it is building upon their VALMOD paper and it may be useful for your PR

[mihailescum]

Hi! I haven't seen it, thank you, I'll look at it.

[hari-packetai]

@mexxexx Did you happen to see this recent paper by Keogh's group on motif/discord discovery? It looks like it is building upon their VALMOD paper and it may be useful for your PR

Hi, Are the algorithms for motif and discord discovery described in this paper ("Matrix profile goes MAD") available in STUMPY?

[seanlaw]

Hi, Are the algorithms for motif and discord discovery described in this paper ("Matrix profile goes MAD") available in STUMPY?

@hari-packetai I do not believe so. The only methods that we currently have are motifs and match. Neither account for variable length matrix profiles