Tag: machine learning

Automating the voting system

MetaMetaData

For the last several years, one of the things our community has struggled with is a lack of active voters. We’ve tried to implement various measures to decrease the need for voters and load for the wonderful ones that actually do actively look through edits and help vote on them—e.g., making more edits auto‐edits and decreasing amount of time edits stay open. However, the edit queue is still quite unwieldy and as such we’ve kept trying to come up with other ways to decrease the load on our contributors.

Over the past few months since our last summit, we’ve been working on training AIs, both for recommendation engines and data analytics, and for helping out with spam, but it soon appeared that we had another valuable dataset: our history of 15,693,824 votes from 16,336 voters and 56,374,198 edits from 2,007,134 editors. It turns out that this is an unintended side-effects of the editing and voting system in that it creates a paper trail of our habits as a community and our collective mind.

A paper trail that you could, say, train a neural network on. And that’s just what we did.

By feeding data from our top voters, we’ve been able to train our network to replicate with 96.4% accuracy the personality when using the other half as test data. That figure is the average for 300 bots each based on our top 300 voters.
We were really impressed with the results but the story doesn’t stop there…

Meet BrainzVoter

The next logical step was to create our own Frankenstein’s monster. By training on 70% of our entire set of votes, we gave birth to a voting bot that represents the essence of our community. “BrainzVoter”, as we dubbed it, is precise and scores a staggering 98.9% accuracy on test data and comparing with the other 30% of our dataset.

To quote the late Terry Pratchet:

Ankh-Morpork had dallied with many forms of government and had ended up with that form of democracy known as One Man, One Vote. The Patrician was the Man; he had the Vote.

Edit filters

In view of the recent developments on net neutrality taken by the European Union with articles 11 & 13/17, MusicBrainz is taking measures to protect against copyright infringement: we’re implementing automatic edit filters. BrainzVoter will use the latest in NLP technology to understand what you, the editors, write in your edit notes, and use this understanding to vote on your edit. It will also inspect any URLs included in the edit note to cross-reference the data. The aggregate data will not be available to the public.

Edits with better and clearer notes will become more likely to pass. Consider this a good opportunity to (re‐)read How to Write Edit Notes!

How will this affect me as an editor?

Not much will change, and you can continue doing what you were doing before! We recommend that you take the time to make clear statements in your edit notes.
You will also be able to use a system of tags to express intent, using for example #typo #correction in the content of your edit text. Syntax highlighting and shortcuts will be available in the text editor.

In the end, by removing the need for humans to look over edits, the bot should give you, the editor, more time to add and edit and fix data in MusicBrainz, without having to spend time checking everyone else’s edits or worry about other editors disagreeing with yours!

After a brief trial period on MusicBrainz, this system will be adapted and also rolled out to BookBrainz.

We hope you will share our excitement for the benefits of automation and help us improve our training models over time. I, for one, welcome our AI overlords.

AcousticBrainz at the 2018 MetaBrainz Summit

We had an in-person meeting at the MTG during the MetaBrainz summit to discuss the status and future of AcousticBrainz. We came up with a rough outline of things that we want to work on over the next year or so. This is a small list of tasks that we think will have a good impact on the image of AcousticBrainz and encourage people to use our data more.

State of AcousticBrainz

AcousticBrainz has a huge database of submissions (over 10 million now, thanks everyone!), but we are currently not using the wealth of data to our advantage. For the last year we’ve not had a core developer from MetaBrainz or MTG working on existing or new features in AcousticBrainz. However, we now have:

  • Param, who is including AcousticBrainz in his role with MetaBrainz
  • Rashi, who worked on AcousticBrainz for GSoC and is going to continue working with us
  • Philip, who is starting a PhD at MTG, focused on some of the algorithms/data going into AcousticBrainz
  • Alastair, who now has more time to put towards management of the project

Because of this, we’re glad to present an outline of our next tasks for AcousticBrainz:

Short-term

Some small tasks that are quick to finish and we can use to show off uses of the data in AcousticBrainz

Merge Philip’s similarity, including an API endpoint

Philip’s masters thesis project from last year uses PostgreSQL search to find acoustically similar recordings to a target recording. This uses the features in AcousticBrainz. We need to ensure that PostgreSQL can handle the scale of data that we have.

An extension of this work is to use the similarity to allow us to remove bad duplicate submissions (we can take all recordings with the same MBID and see if they are similar to each other, if one is not similar we can assume that it’s not actually the same as the other duplicates, and mark it as bad). We want to make these results available via an API too, so that others can check this information as well.

Merge Existing PRs

We have many great PRs from various people which Alastair didn’t merge over the last year. We’re going to spend some time getting these patches merged to show that we’re open to contributions!

Publish our Existing models

In research at MTG we’ve come up with a few more detailed genre models based on tag/genre data that we’ve collected from a number of sources. We believe that these models can be more useful that the current genre models that we have. The AcousticBrainz infrastructure supports adding new models easily, so we should spend some time integrating these. There are a few tasks that need to be done to make sure that these work

  • Ensure that high-level dumps will dump this new data (If we have an existing high-level dump we need to make a new one including the new data)
  • Ensure that we compute high-level data for all old submissions (we currently don’t have a system to go back and compute high-level data for old submissions with a new model, the high-level extractor has to be improved to support this)

Update/fix some pages

We have a number of issues reported about unclear text on some pages and grammar that we can improve. Especially important are

  • API description (we should remove the documentation from the main website and just have a link to the ReadTheDocs page)
  • Front page (Show off what we have in the project in more detail, instead of just a wall of text)
  • Data page (instead of just showing tables of data, try and work out a better way of presenting the information that we have)

Fix Picard plugin

When AB was down during our migration we were serving HTML from our API pages, which caused Picard to crash if the AB plugin was enabled while trying to get AB data. This should be an easy fix in the Picard plugin.

High Impact

These are tasks that we want to complete first, that we know will have a high impact on the quality of the data that we produce.

Frame-level data

We want to extract and store more detailed information about our recordings. This relies on working being done in MTG to develop a new extractor to allow us to get more detailed information. It will also give us other improvements to data that we have in AB that we know is bad. This data is much bigger than our current data when stored in JSON (hundreds of times larger), so we need to develop a more efficient way of storing submissions. This could involve storing the data in a well-known binary data exchange format. A bunch of subtasks for this project:

  • Finish the essentia extractor software
  • Decide on how to store items on the server (file format, store on disk instead of database)
  • Work out a way to deal with features from two versions of the extractor (do we keep accepting old data? What happens if someone requests data for a recording for which we have the old extractor data but not the new one?)
  • Upgrade clients to support this (Change to HTTPS, change to the new API URL structure, ensure that clients check before submission if they’re the latest version, work out how to compress data or perform a duplicate check before submission)
  • Deduplication (If we have much larger data files, don’t bother storing 200 copies for a single Beatles song if we find that we already have 5-10 submissions that are all the same)

MusicBrainz Metadata

Rashi’s GSoC project in 2018 helped us to replicate parts of the MusicBrainz database into AcousticBrainz. This allows us to do amazing things like keep up-to-date information about MBID redirects, and do search/browse/filtering of data based on relationships such as Artists just by making a simple database query. We want to merge this work and start using it.

Dumps

When we changed the database architecture of AcousticBrainz in 2015 we stopped making data dumps, making people rely on using the API to retrieve data. This is not scalable, and many people have asked for this data. We want to fix all of the outstanding issues that we’ve found in the current dumps system and start producing periodic dumps for people to download.

Build more models

In addition to the existing models that we’ve already built (see above, “Publish our Existing models”), we have been collecting a lot of metadata that we could use to make even more high-level models which we think will have a value in the community. Build these models and publicly release them, using our current machine learning framework.

Wishlist

These are tasks that we want to complete that will show off the data that we have in AcousticBrainz and allow us to do more things with the data, but should come after the high-impact tasks.

Expose AB data on MusicBrainz

As part of the process to cross-pollinate the brainz’s, we want to be able to show a small subset of AB data that we trust on the MB website. This could include information such as BPM, Key, and results from some of our high-level models.

Improve music playback

On the detail page for recordings we currently have a simple YouTube player which tries to find a recording by doing text search. We want to improve the reliability and functionality of this player to include other playback services and take advantage of metadata that we already have in the MusicBrainz database.

Scikit-learn models

The future of machine learning is moving towards deep learning, and our current high-level infrastructure written in the custom Gaia project by MTG is preventing us from integrating improved machine learning algorithms to the data that we have. We would like to rewrite the training/evaluation process using scikit-learn, which is a well known Python library for general machine learning tasks. This will make it easier for us to take advantage of improvements in machine learning, and also make our environment more approachable to people outside the MusicBrainz community.

Dataset editor improvements

Part of the high-level/machine learning process involves making datasets that can be used to train models. We have a basic tool for building datasets, however it is difficult to use for making large datasets. We should look into ways of making this tool more useful for people who want to contribute datasets to AcousticBrainz.

Search

With the integration of the MusicBrainz database into AcousticBrainz, we will be able to let people search for metadata related to items which we know only exist in AcousticBrainz. We think that this is a good way for people to explore the data, and also for people to make new datasets (see above). We also want to provide a way that lets people search for feature data in the database (e.g. “all recordings in the key of Am, between 100 and 110BPM”).

API updates

As part of the 2018 MetaBrainz summit we decided to unify the structure of the APIs, including root path and versioning. We should make AcousticBrainz follow this common plan, while also supporting clients who still access the current API.

We should become more in-line with the MetaBrainz policy of API access, including user-agent reporting, rate limiting, and API key use.

Request specific data

Many services who use the API only need a very small bit of information from a specific recording, and so it’s often not efficient to return the entire low-level or high-level JSON document. It would be nice for clients to be able to request a specific field(s) for a recording. This ties in with the “Expose AcousticBrainz data on MusicBrainz” task above.

Everything else

Fix all our bugs and make AcousticBrainz an amazing open tool for MIR research.

Thanks for reading! If you have any ideas or requests for us to work on next please leave a comment here or on the forums.

GSoC 2018: SpamBrainz – Fighting spam in MusicBrainz using machine learning

Hi, I’m Leo and I spent my summer building and training SpamBrainz, our new solution to fighting spam in MusicBrainz. If you haven’t heard of SpamBrainz before it’s probably because it did not exist before this year’s Summer of Code.

For quite a while now the amount of spam in MusicBrainz has started to become a serious problem. Often this means editors are automatically created with descriptions that look not unlike the spam emails most of us get every day, promoting other websites and services.

During last year’s MetaBrainz Summit we discussed possible solutions to this and came up with the Spam Ninja system. Essentially this means that Soon™ there will be a group of editors that receive spam reports and have the ability to delete editors and entities that are nothing but spam.

Now with MusicBrainz having almost two million registered editors, could we really expect the Spam Ninjas to manually check every single one of them in addition to all the new registrations? Obviously not, and this is where SpamBrainz comes in.

SpamBrainz is a machine learning system that looks at all editors and decides whether or not it thinks they are spammers. If it thinks they are, it automatically notifies the spam ninjas who then decide whether or not SpamBrainz was correct.

What’s great about this system is that a human is guaranteed to look at any report and at no point does a computer decide that you’re a spammer and should be banned, because no one wants machines to run the world, right?

Building SpamBrainz

While most GSoC projects involve adding features to existing systems, SpamBrainz is something entirely new and I had not built anything on this scale before so I started out by doing tons of research.

When building a machine learning project you should always start by doing some good
old statistics first and trying to figure out what matters about your data and how the
system could use it. I wrote a couple Jupyter notebooks (which are great for working with data) to do this.

As I was not working for MetaBrainz at the time and had to respect our privacy policy, I wrote a script to collect the most common values of a couple different editors, anonymize them and save them to a report. Using that data I could compare all spam and non-spam editors and decide upon a set of datapoints that would be useful for my machine learning model. Yvanzo then ran these on the live database and I could happily do my data analysis without compromising user privacy.

Next I built a pretty boring Flask-based API that would allow MusicBrainz to queue up editor analysis and training. Quite a few different MetaBrainz projects use Python and need to access the MusicBrainz database so a long time ago someone wise decided to move commonly used code into a repository called brainzutils-python. All I had to do was to add some code for accessing editor data through it.

In a surprise move by ruaok I was then hired by MetaBrainz as a contractor with a yearly salary of 100g of chocolate. I probably should have negotiated what kind of chocolate but what mattered most was that I could now work with user data without breaching our privacy policy.

But before I could build my Keras model I had to decide on a final set of input features and do write code for preprocessing the data. Only then could I finally get started building and testing models.

The current SpamBrainz state of the art model is Lodbrok which actually turned out to work really well, reaching a 99% accuracy in detecting spam while only mis‐classifying 0.2% of real users as spammers. Obviously the latter won’t be a problem because after all a Spam Ninja will still check these reports.

Future outlook

Now that GSoC is over I could just disappear with all the money and leave SpamBrainz in its current state but obviously that’s not what I am planning to do.

I would like to work with zas on getting it deployed along with the Spam Ninja system, improve the code documentation and try to tackle the remaining problem that is online learning (which as it turns out, isn’t as easy as I had thought).

With spam always evolving and spammers already moving to more sophisticated methods than just using editor biographies, I’d also look into building separate models for other entities.

After all SpamBrainz is just getting started and I’m very much looking forward to continuing our journey towards reducing the spam we all have to endure on MusicBrainz and other MetaBrainz projects.