As most of my side projects, “seevl DJ” started as a quick hack on a sunday afternoon. Yet, it has been quickly picked and featured on Fast Company and Hypebot, and also got some attention on Twitter itself.
With a little help from my friends
I’ve spend some time improving it so you can use additional commands, e.g. “a song by”, “play me something like”. In addition, it now uses the Freebase / YouTube mappings combined with the seevl API in order to find an artist’s videos (when using a genre / label / related query).
Last but not least, you can now use “/cc @user” and “for @user” in your Tweet to send a track to any of your friend, the music video being available directly on their feed through Twitter cards (Web and mobile).
Thinking again about Twitter as an intelligent agent on the Web, let’s be bold and imagine this integrated with the buy / Stripe integration. While it’s now used to buy stuff, what about paying for services with it? “Hey @uber, bring @myfriend here”. “Hey @trycaviar, sushis for 6 please”. Both answering with an automated tweet embedding a Buy button so you can validate the order; and get your black car or food home within minutes. All through Twitter.
Natural Language Processing is one way to enable this, but another one is to pre-fill such “service-based tweets” so that users would just have to complete a few fields (e.g. number of people when messaging @opentable). This makes things much easier from the processing side, also providing a friction-less experience to users. Technically, the intelligence can be brought by schema.org actions, as I’ve wrote in the past, using JSON-LD as the supporting data serialisation.
Note that it might conflict with the previous Twitter TOS if you unfollow too many people at once. However, it will happen only once if you put it into a daily crontab. It was safe in my case, but I can’t guarantee it will be in yours. You may also reach the API rate-limiting if you’ve too many followees.
It’t built using python-twitter, and is available under the MIT license.
While the Twitter music app eventually failed, it’s still clear that people use Twitter’s data stream to share and/or discover new #music. Thanks to Twitter cards, a great thing is that you can directly watch a YouTube video, or listen to a SoundCloud clip, right from your feed, without leaving the platform. But what if Twitter could be your own DJ, playing songs on your request?
Since it’s been a few month since I enjoyed my last Music Hack Day – oh, I definitely miss that! – I’ve hacked a proof of concept using the seevl API, combined with the Twitter and the YouTube ones, to make Twitter acts as your own personal DJ.
Hey @seevl, play something cool
The result is a twitter bot, running under our @seevl handle, which accepts a few (controlled) natural-language queries and replies with an appropriate track, embedded in a Tweet via a YouTube card. Here are a few patterns you can use:
Hey @seevl, play something like A
To play something that is similar to A. For instance, tweet “play something like New Order”, and you might get a reply with a Joy Division track in your feed.
Hey @seevl, play something from L
To play something from an artist signed on label L (or, at least, that used to be on this label at some stage)
Hey @seevl, play some G
To play something from a given genre G
Hey @seevl, play A
To simply play a track from A.
By the way, you can replace “Hey” by anything you want, as long as you politely ask your DJ what you want him to spin. Here’s an example, with my tweet just posted (top of the timeline), and a reply from the bot (bottom left).
A little less conversation
As it’s all Twitter-based, not only you can send messages, but you can have a conversation with your virtual DJ. Here’s for instance what I’ve sent first
It’s kind of fun, I have to say, especially due to the instantaneous nature of the conversation – and it even reminds IRC bots!
Unfortunately, it’s likely that the bot will reach the API rate-limit when posting Tweets (and I’m not handling those errors in the current MVP), so you may not have a reply when you interact with it.
Twitter As A Service?
Besides the music-related hack, I also wanted to showcase the growth of intelligent services on the Web – and how a platform like Twitter can be part of it, using “Twitter As A Service” as a layer for an intelligent Web.
The recently-launched “Buy button” is a simple example of how Twitter can be a Siri-like interface to the world. But why not bringing more intelligence into Twitter. What about “Hey @uber, pick me in 10 minutes”, and using the Tweet geolocation plus a Uber-API integration integration to directly pick – and bill – whoever #requested a black car? Or “Please @opentable, I’d love to have sushis tonight”, and get a reply with links to the top-rated places nearby, with in-tweet booking capability (via the previous buy button)? The data is there, the tools and APIs are there, so…
More Products: A Chrome plug-in for Product Hunt recommendations
With more than 6,000 products already in the Product Hunt database, I’ve decided to use the API to build a product recommendation engine. It seems that evey times it comes to hacking and APIs, I can’t get away of discovery, or music. Or both.
The result is a Chrome extension simply named “More Products!”. It directly integrates top-10 related products for each product page, as you can see below. I might iterate on the algorithm itself, but want to keep this plug-in very focused so it’s unlikely that it will integrate other features. Note that it doesn’t track anything, so your privacy is preserved.
Under the hood
The engine relies on the API to get the list of all products and related posts, and then uses TF-IDF and Cosine similarity to to find similarities between them, using NLTK and scikit-learn, respectively the standard Python tools for Natural Language Processing and Machine Learning. To put it simply, it builds a giant database of words used in all posts, mapped to products with their frequency, and then finds how close products are, based on those frequencies.
New products are fetched every 2 hours, and recommendations are updated at the same time. Flask handles requests between the extension and the recommendations database, and Redis is used as a cache layer.
Here’s the second post of my data analysis series on the Rolling Stone top 500 greatest songs of all time. While the first one focused on lyrics, this one is all about the acoustic properties of the data-set – especially their volume and tempo.
To do so, I used the EchoNest, which delivers a good understanding of each track at the section level (e.g. verse, chorus, etc.) but also at a deeper “segment” level, providing loudness details about very short intervals (up to less than a second). This is not perfect, due to some issues discussed below, but gives a few interesting insights.
[Update 22 July] As noted in the comments, there were a few unexpected results. I’ve run the pipeline again and done some more cleaning on the API results, as explained here.
Black leather, knee-hole pants, can’t play no highschool dance
As my goal was to identify relevant tracks from the dataset, in addition to absolute values for the loudness and tempo of each track, I also looked at their standard deviation. If you’re not familiar with it, this helps to identify which songs / artists tend to be closer to their average tempo / loudness, versus the ones that are more dynamic.
Before going through individual songs from the top-500, let’s take an example with the top-10 Spotify tracks of a few artists to check their loudness:
And the tempo:
You can see that some bands really deserve their reputation. For instance, while the Pink Floyd have a high standard deviation both in volume and tempo (not surprising), Motörhead is not only the loudest (in average) of the list, but also the one with the smallest standard deviation, meaning most of their tracks evolve around that average loudness. In order words, they play everything loud. While the Ramones and just fast, everything fast. And when they’re together on stage, the result is not surprising
But you don’t really care for music, do you?
Coming back to the top 500, I ran the Echonest analysis on 474 tracks of the list. The 26 missing are due to various errors at different stages of the full pipeline.
On the one hand, I’ve used raw results from the song API to get the average values. [Update 22 July] I had to consolidate the data by aggregating multiple API results together. For a single song, multiple tracks are returned by the API (as expected), but there can be large inconsistencies between them. For instance, if you search for American Idiot, one track (ID=SOHDHEA1391229C0EF) is identified having a tempo of 93, the other one (SOCVQDB129F08211FC) of 186. Some can also have slighter variations (in volume for instance, between a live and the original version). To simplify things – and I agree it include a bias in the results – I averaged the first 3 results from the API.
On the other hand, I relied on numpy to compute the standard deviation from the first API result, removing first the fade-in and fade-out of each track. Here, I’ve also skipped every segment of section where the API confidence was too low (< 0.4).
I’m waiting for that final moment you say the words that I can’t say
Last but not least, I’ve normalized and combined both the tempo deviation and the rhythm one to assign a [0:1] score to each track in order find the most and less dynamic tracks overall. Here’s the top-5 of the most dynamic ones:
If you listen to My Generation, you can clearly hear the dynamic both in tempo and volume with the different bursts of the song. While the Radiohead one is more on the long-run, with clearly distinct phases as shown below for the volume part.
Finally, here are the less dynamic ones. Several ones on that list made it through the charts, showing that even though a song can be pretty flat in both volume and tempo, it can still be a hit – or at least an earworm:
While toying with the public BigQuery datasets, impatiently waiting for Google Cloud Dataflow to be released, I’ve noticed the Wikipedia Revision History one, which contains a list of 314M Wikipedia edits, up to 2010. In the spirit of Amazon’s “people who bought this”, I’ve decided to run a small experiment about music recommendations based on Wikipedia edits. The results are not perfect, but provide some insights that could be used to bootstrap a recommendation platform.
Here, my assumption to build a recommendation system is that Wikipedia contributors edit similar pages, because they have an expertise and interest in a particular domain, and tend to focus on those. This obviously becomes more relevant at the macro-level, taking a large number of edits into account.
In the music-discovery context, this means that if 200 of the Wikipedia editors contributing to the Weezer page also edited the Rivers Cuomo one, well, there might be something in common between both.
This dataset contains a version of that data from April, 2010. This dataset does not contain the full text of the revisions, but rather just the meta information about the revisions, including things like language, timestamp, article and the like.
Number of rows: 314M
Sounds not too bad, as it contains a large set of (page, title, user) tuples, exactly what we need to experiment.
Querying for similarity
Instead of building a complete user/edits matrix to compute the cosine distance between pages, or using a more advanced algorithm like Slope One (with the number of edits as an equivalent for ratings), I’m simply finding common edits, as explained in the original Amazon paper. And, to make this a bit more fun, I’ve decided to do it with a single query over the 314M rows, testing BigQuery capabilities at the same time.
The following query is used to find all pages sharing common editors with the current ones, ranked by the number of common edits. Tested with multiple inputs, it took an average of 5 seconds to answer it over the full dataset. You can run those by yourself by going to your Google BigQuery console and selecting the Wikipedia dataset.
SELECT title, id, count(id) as edits
WHERE contributor_id IN (
AND contributor_id IS NOT NULL
AND is_bot is NULL
AND is_minor is NULL
AND wp_namespace = 0
GROUP BY contributor_id
AND is_minor is NULL
AND wp_namespace = 0
GROUP EACH BY title, id
ORDER BY edits DESC
This is actually a simple query, finding all pages (wp_namespace=0 restricts to content pages, excluding user pages, talks, etc.) edited (excluding minor edits) by users whom also edited (excluding bots and minor contributions) the page with ID 30423, ranking them by number of edits. You can read it as “Find all pages edited by people who also edited the page about the Clash, ranked by most edited first”.
And here are some of the results
As you can see, from a music-discovery perspective, that’s a mix between relevant ones (Ramones, Sex Pistols), and WTF-ones (The Beatles, U2). There’s also a need to exclude non-music pages, but that could be done programmatically with some more information in the dataset.
Towards long tail discovery
As we can expect, and as seen before, results are not that good for mainstream / popular artists. Indeed, edits about the Beatles page are unlikely, in average, to say much about the musical preferences of their editors. Yet, this becomes more relevant for long-tail artist discovery: if you care editing indie bands pages, that’s most likely you care about it.
Trying with Mr Bungle, the query returns Meshuggah and The Mars Volta as the first two music-related entries, all of them playing some kind of experimental metal – but then digresses again with the Pixies. Looking at band members / solo artists and using Frank Black as a seed leads to The Smashing Pumpkins, Pearl Jam, R.E.M. and obviously the Pixies as the first four recommendations. Not perfect for both, but not too bad for an algorithm that is completely music-agnostic!
Scaling the approach
There are many ways this could be improved, for instance:
Removing too-active contributors – who may edit pages to ensure Wikipedia guidelines are followed, rather than for topic-based interest, and consequently introduce some bias;
Filtering the results using some ML approaches or graph-based heuristics – e.g. exclude results if their genres are more than X nodes away in a genre taxonomy.
Using time-decay – someone editing Nirvana pages in 1992 might be interested in completely different genres now, so joint edits might not be relevant if done with an x-days interval or more.
Yet, besides its scientific interest, and showing that BigQuery is very cool to use, this approach also showcases – if needed – that even though algorithms may rule the world of music discovery, they might not be able to do much without user-generated content.