glenn mcdonald furia furialog · The War Against Silence · photography · songs · code · other things     ↑vF
29 September 2009 to 25 February 2009
Stefano Mazzocchi (of Freebase) and David Karger (of MIT) have been holding a slow but interesting conversation about data reconciliation. It's been phrased as a sort of debate between two arbitrarily polarized approaches to the problem of cleaning up data so that you don't have multiple variant references to the same real data-point: either you try to do this cleanup "in the data", so that it's done and you don't have to worry about it any more; or you leave the data alone and figure you'll incorporate the cleanup into your query process when you actually start asking specific questions.  

But I think this is less of a debate than Stefano and (particularly) David are making it seem. Or, rather, that the real polarization is along a slightly different axis. The biggest difference between Stefano's world and David's happens before you even start to worry about when you're going to start worrying about data cleanup: Freebase is attempting to build a single huge structured-data-representation of (eventually) all knowledge; David's work is largely about building lots of little structured-data-representations of individual (relatively) small blobs of knowledge.  

Within a dataset, though, Stefano and David (and I) are in enthusiastic agreement: internal consistency is what makes data meaningful. If your data doesn't know whether "Beatles" and "The Beatles" are one band or two different ones, your answers to any question that touches either one are liable to be so pathetically wrong that people will very quickly stop asking you things. In the classic progression from Data to Information to Knowledge to Wisdom, this is what that first step means: Data is isolated, Information is consolidated. It would be inane to be against this. (Unless, I guess, you were also against Knowledge and Wisdom.)  

It's on the definition of "within", though, that most of the interesting issues under the heading of "the Semantic Web" wobble. David says "I argued in favor of letting individuals make their own [datasets] ... and worry about merging them with other people’s data later." He really does mean individuals, and the MIT-developed tool he has in mind (Exhibit) is designed (and only suited) for fairly small datasets. Freebase is officially a database of everything, but of course within this everything are lots of subsets, and Stefano is really talking more about cleaning up these subsets than the whole universe. Reconciling all the artist references to "Beatles" and "The Beatles" from albums and songs is a straightforward practical task both beneficial and probably necessary for asking questions of a song/album/artist dataset, whether it's embedded in a larger one or not. Reconciling the "'The Beatles" who are the artist of the song "Day Tripper" with "The Beatles" who are depicted in cartoon form on a particular vintage lunchbox for sale on eBay, on the other hand, is less conceptually straightforward, and of more obscure practical import.  

The thing I'm working on at ITA falls in between Exhibit and Freebase on this dataset axis, both in capacity and design. We handle datasets much bigger than you could put in Exhibit, and allow you to explore and analyze them in ways that Exhibit cannot; but we separate datasets, partly for scalability but even more importantly to specifically keep out of any unnecessary quagmire of trying to reconcile not just the bits of data but the structures.  

And my own scouting report, from having done a lot of data-reconciliation in a system designed for it, and in other lives of a lot of more-painful data-reconciliation in various systems not really designed for it, is the same as Stefano's report from his experiences at Freebase: getting from data to information, at scale, is hard. It's mainly hard in ways that machines cannot just solve for you, and anybody who thinks RDF is a solution to data-reconciliation is, as Stefano puts it, confusing model mixability with semantic reconciliation, and has probably not noticed because they've only been playing with toy datasets, which is like thinking you're learning Architecture by building a few birdhouses.  

And this is all exactly why I have repeatedly argued for treating the "Semantic Web" technology problem as a database-tech problem first, and a web-tech problem only secondarily. David complains, justifiably, about the pain of combining two small, simple folk-dance datasets he himself cares about. But just as Stefano says in another example, the syntax problems here (e.g. text in a YouTube description box, rather than formally-labeled data records) are fairly trivial compared to the modeling differences. All the URIs and XSL transformations in the world are not going to allow every two datasets to magically operate as one. Some person, without or without tools to magnify their effort, is going to have to rephrase one dataset in the argot of the other.  

And to say another thing I've said before again, the fact that RDF doesn't itself solve these problems is not its terminal design flaw. Its terrible flaw is that it isn't even a sufficient foundation upon which to build the tools that would help solve these problems. It takes the right fundamental idea, that the most basic conceptual structure of data is things and the relationships between them, but then becomes so invertedly obsessed with the theoretical purity of this reduction that it leaves a whole layer of needed practical elaboration unbuilt. We shouldn't need abstruse design patterns to get ordered lists, or rickety reasoning engines just to get relationships that go both directions, or endless syntax officiousness that gets us expensive precision with no gain in accuracy.  

This effort has let us down. And worse, it has constrained smart people so that their earnest efforts cannot help but let us down. After years and years of waiting, we have no Semantic Web of information, we have only Linked Data, where the word "Linked" is so tenuously justified it might as well be replaced with some pink-ink-drinking Seuss pet, and the word "Data" is tragically all too accurate. We have all these computers, living abbreviated lives of quiet depreciation, filled with the data that should become our wisdom, and yearning, if they are allowed to yearn for even one thing, to be able to tell us what they know.
( ) Eight Lines of Estrangement  

[Fair warning: This is another post about data-modeling and query languages, and it isn't likely to be the last. It may or may not be interesting to people with personal interests in those topics, but I think it's pretty unlikely to be interesting to people without those interests. You have been warned.]  
 

In data-modeling you usually live in fear of the word "usually". Accounting for the subsequent "but sometimes" is usually where a simple, manageable data-model starts its ugly metamorphosis towards tangled and unusable. Same with "mostly", "more often", "occasionally". Most data-modeling contexts are only really ever happy with "always" and "never"; and the real world is usually not so helpfully binary.  

DiscO, my data-model for discographies, notes that Tracks, Releases and Sequences can all have Artists, but that usually Tracks would get their Artist indirectly from their Release, which would in turn get it from its Sequence.  

What this means, in practical terms, is that when we're entering or importing data, we don't necessarily want to have to set Artist individually on every single Sequence, Release and Track. But when we're querying the data, we want to be able to ask about Artist on every one of those.  

Data-modeling people call this "inference", and it's a whole academic subject of its own, deeply entangled with abstract logic and belief-system consistency. The Sequence/Release/Track problem sounds theoretically straightforward at first, but gets very hard very quickly once you realize that some Releases are compilations of Tracks by multiple artists, and some Sequences have Releases by multiple artists. Thus it's not quite true that Sequence "contains" Release, and upon that "not quite" most academic approaches will founder and demand to be released from this vagueness via a rococo expansion of the domain model to separate the notions of Single-Artist and Multiple-Artist Sequences and Releases.  

But "usually" is a good concept. And for lots of real-world data problems it can be handled without flailing into existential abstraction. We can keep our model simple, and fill in the implied data with a very general mechanism: the relationships between "actual" values and "effective" values can themselves be described in queries.  

Releases, we said, may have Artists directly, or may get them implicitly from their Sequence. More specifically, we probably mean something like this:  

- If a Release has an Artist, directly, use that.
- If it doesn't, get all the Sequences in which it occurs.
- Ignore any Sequence that itself has no Artist or multiple Artists.
- If all the remaining (single-Artist) Sequences have the same Artist, use that.
- Otherwise we don't know.  

This can be written out in Thread in pretty much these exact steps:  

Release|Artist=(.Artist;(.Sequence:(.Artist::#=1).Artist::#=1))
 

This isn't a syntax tutorial, but ";" means otherwise, and "::#=1" tests a list to see if it has exactly one entry, so maybe you can sort of see what the query is doing.  

A Track, then, follows the same pattern, but has to check one more level:  

- If a Track has an Artist, directly, use that.
- If it doesn't, get all the Releases in which it occurs.
- Ignore any Release that itself has no Artist or multiple Artists.
- If all the remaining (single-Artist) Releases have the same Artist, use that.
- Otherwise, get all the Sequences for all the Releases on which the track appears.
- Ignore any Sequence that itself has no Artist or multiple Artists.
- If all the remaining (single-Artist) Sequences have the same Artist, use that.
- Otherwise we don't know.  

Or, in Thread:  

Track|Artist=(
.Artist;
.(.Release:(.Artist::#=1).Artist::#=1);
.(.Release.Sequence:(.Artist::#=1).Artist::#=1))
 

In this system, once you've learned the query-language (and I'm not saying this is a trivial task, but it's not that hard), you can do most anything. The language with which you ask questions is the same language you use for stipulating answers.  
 

Query-language-geek postscript: And, of course, it's the same language you use for obsessively fiddling with your questions and answers because you just can't help it. That Track query has some redundancy, and while redundancy isn't always bad, it's almost always fun to see if you can get rid of it. In this case we're asking the same question ("What's your artist?") in all three steps of the query. We can rearrange it so that we get all the Tracks and Releases and Sequences first, then ask all the Artist questions at the end:  

Track|Artist=(...__,Release,Sequence:(.Artist::#=1).Artist::#=1)
 

"...__," may initially look a little like ants playing Limbo, but "...__,Release,Sequence" means "get these things, their Releases and their Sequences, and then add those things' Releases and Sequences, etc. until you get to the end. So this version of the query builds up the complete list of all the places from which this Track could derive its artist, keeps only the ones that have a single artist, and then sees if we're left with a single artist at the end of the whole process. Depending on what we want to do if our data has internal contradictions, this might actually be functionally better than the first version, not just shorter to type.  

But DiscO also has all that stuff about Original Versions, and it would be nice if our Artist inference used that information, too. If Past Masters [Remastered] is an alternate version of Past Masters, Vol. 1 & 2, and Past Masters, Vol. 1 & 2 belongs to the Beatles' Compilations sequence, then we should be able to deduce that the version of "We Can Work It Out" on Past Masters [Remastered] is by the Beatles. Our experimental elimination of redundancy now pays off a little bit, because we only have to add in "Original Version" in one place:  

Track|Artist=(...__,Release,Sequence,Original Version:(.Artist::#=1).Artist::#=1)
 

And interestingly, since both Tracks and Releases can have Original Versions, this whole thing actually works for either type, and thus we can combine the two and have even fewer things to worry about:  

Track,Release|Artist=(...__,Release,Sequence,Original Version:(.Artist::#=1).Artist::#=1)
 

Having fewer things to worry about is (usually) good.
What Beatles album is "Day Tripper" on?  

This is partially a trick question, of course, as "Day Tripper" was originally a non-album single, but it has been on several Beatles compilations over the years, including the red 1962-1966 best-of, and in the remastered 2009 catalog it lands on both the mono and stereo versions of Past Masters.  

Starting from scratch, it would take you, as a person, a little while to figure this answer out on the web. There's a Wikipedia page for the song, which is the top search hit for the above question in both Google and Bing, and it explains the non-album-ness and mentions several compilations, but it doesn't (at least as of this moment) clarify current availability, and none of the pages for the non-current compilations refer you explicitly (again, as of this moment) to the right place.  

There are a few paths that lead you to a voluminous page for the Beatles discography, on which "Day Tripper" is again mentioned several times, but this page doesn't itemize the track-listings of compilations, and the "Day Tripper" links just go back to the original song-page. But eventually you might blunder into the separate pages for the Mono and Stereo box sets, and from there you might wander over to Amazon.  

Here even more potential confusion awaits you. The top Google hit for "amazon past masters", at the moment, is the now-obsolete second volume of the old CD edition. Searching on "past masters" on Amazon itself gets you the Remastered edition as the top hit, but idiotically suggests buying it together with the old Volume 2, and you would have to scrutinize the track lists to realize that Past Masters [Remastered] subsumes Past Masters, Vol. 1 and Past Masters, Vol. 2.  

In fact, if you backtrack and try searching for "Day Tripper" on Amazon, the top hit is Rubber Soul, the album on which "Day Tripper" would have appeared, chronologically, but doesn't. And, for good measure, that top hit is the obsolete 1990 CD, not the new remaster.  

Bleh.  
 

But at least you're a person, and via a judicious combination of intuition and stubbornness and asking some friends who might know, you can eventually solve these information problems. If you were a computer, you'd be fucked.  

Which is OK on some existential level, because if you were a computer you probably wouldn't appreciate the song, anyway. But the point of computers is to help people do things, and a computer ought to be a particularly helpful tool when the thing you need to do is sort through some data.  

But for the computer to help you puzzle through this data, the data has to be modeled usefully by people first. There are several prominent sources of meticulously structured data about music, so this should be easy. But here, sadly, people have let us down again. And again, and again. Let's see how.  
 

All Music Guide  

A text-search for "Day Tripper" (there's no other query interface) returns a full page of cryptic results. There's an "Occurrences" column, and although it's not clear exactly what that means, it's obvious that more is supposed to be better, and the first listing has 360 where none of the rest have more than 13, so presumably that's the "right" one.  

Clicking this gets you 8 pages of results, which is annoying in itself (the splitting of them into pages, I mean). They're sorted by Artist, which sounds reasonable enough, except that the ones without artists are sorted first, and thus the first page of results is almost totally crap. There are lots of Beatles releases listed, but they get split between pages 1 and 2 of the results list, making it impossible to look at them all at once.  

But if you drill into one of them at random, and then click on "Day Tripper" in the track listing, you do finally get to a page that lists all (or several, anyway) Beatles releases on which this song appears. There are 24, though, including such things as "Five Nights in a Judo Arena", which human intuition might guess is not a normative release, but a computer would have no basis for dismissing. These releases are in date-order, at least, but this turns out to be worse than worthless for our current question, because All Music has modeled Past Masters [Remastered] not as an album, but as an alternate manifestation of the album Past Masters, Vol. 1 & 2, which means it appears way up in the middle of this list, labeled 1988, because that was the year of its earliest issue (on cassette!).  

Looking through the data, in fact, we see that although All Music has lots of individual detail on most kinds of things, it has essentially nothing that models the relationships of things to each other, or in groups. There is no modeled connection between Past Masters, Vol. 1 & 2, Past Masters, Vol. 2, The Beatles: Mono Box Set and The Beatles: Stereo Box Set, even though v1/2 subsumes v2, and both boxes subsume both.  

And there's no modeling of "in print", or any notion of representing the subset of albums that represent the current core catalog. So a computer can't use this data to answer real questions by itself. Source fail.  
 

MusicBrainz  

This is a database first, not a guide, and thus a more likely candidate for well-structured data anyway, and one where I won't pick at their explicitly-secondary browsing UI.  

The good news is that MusicBrainz has the kind of data we need. They have some relationships between tracks, like one being a mashup of some others, so presumably they could add one to express that the Mono Masters of "Day Tripper" is a different version from the Past Masters [Remastered] one, but the same underlying song. They already have a reconciliation mechanism by which they can say that the "Day Tripper" on 1962-1966 is "the same" as the one on Past Masters, although at the moment the reconciliation data looks too noisy for real use.  

They even have the notion of one release being part of a set, although I didn't find very many examples of sets, and in particular I can't tell if a release can be part of more than one set. But if they can, that might be a mechanism for expressing official catalogs, current availability, and various other kinds of groupings and subsets.  

So current source fail, but at least there's hope here.  
 

Freebase  

Freebase is easily the most sophisticated public attempt at universal data-modeling, at the moment, but this is a caveat as well as a compliment. Freebase models attempt to represent everything that could possibly exist, and thus tend to drift quickly from usable simplicity towards abstractly-correct awkwardness, usually coming to rest far into the latter.  

So if you search on "Day Tripper", you will find that there are results of that name as "Topic", "Composition", "Musical Album" and "Musical Track", with at least dozens of the latter. Freebase fails the usability test even more spectacularly than All Music, as the list of Musical Tracks is presented with no grouping or distinguishing information at all, just "Day Tripper (Musical Track)" after "Day Tripper (Musical Track)", and you have to click on each one to get any clarifying info. "Day Tripper" the composition does not link to any of the tracks, and "Day Tripper" the album turns out to be a compilation of Beatles covers which does not, at least as far as the listed information shows, even contain the song "Day Tripper".  

And if you delve into the internals of the Freebase music schema, you can quickly develop a guess about why the data has not all been filled in: there's too damn much structure. A music/track is a recording of a music/composition. The track can appear on multiple music/releases, each of which is a publication of a particular music/album. Unless you need to model who was in the band during the making of an album, in which case the album links instead to a set of music/recording_contributions, which is each a combination of albums, contributors and roles.  

Oh, except compositions can be recorded "as albums", in which case they link to music/album without going through music/track, and tracks can appear directly on releases without going through music/album. And there's no current property for saying that a given track is an alternate version of another, but from following Freebase modeling discussions I can confidently guess that they'd model that by saying that a music/track links to something like a music/track_derivation, which itself is a combination of original track, derivative track, deriving artist (or music/track_derivation_contribution) and derivation type. And Freebase's query-language doesn't provide any recursion, so if these relationships chain, good luck querying them.  
 

Music Ontology  

This isn't a database, just an attempt at a model for one. And, grimly, it's another quantum level more elaborately and unusably correct than the Freebase model. Even "MO Basics" (and the "Overview" has 22 more tables of explication beyond these "Basics", without getting into the "details") includes conceptual distinctions between Composition, Arrangement, Recording, Musical Work, Musical Item, MusicalExpression and MusicalManifestation. And then there are pages upon pages of minutely itemized trivia like beginsAtDuration, djmix_of, paid_download (and "paiddownload", which is different in some way I couldn't figure out), AnalogSignal, isFactorOf... This list is bad because it's too long, but the fact that it's in the schema means that it's also bad because no matter how long it is, it will never include every nuance you ever find yourself wanting, and thus over time it will only accumulate more debris.  

A tour-de-force into a cul-de-sac.  
 

The Rest of the Web  

Searching on any particular band or bit of music will unearth dozens or hundreds of other sites that contain bits of the information we need: stores, discographies, databases, forums, fan pages, official sites. Almost universally, these are either unqueriable flat HTML pages, or tree-structured databases with even less interlinking than the above sites. Encyclopaedia Metallum, my favorite metal site, has full track listings for a genuinely mind-boggling number of releases by an astonishing number of bands, but the tracks themselves are not data-objects and a machine can find out nothing about them. There are several lovingly hand-crafted Beatles discographies on the web, all far too detailed for our original casual query, and all essentially useless to a computer attempting to help us.  

So: Ugh. Triple ugh because a) the population of people willing to put time and energy into filling out music-related data-forms is obviously huge, b) the modeling problems are not intractably complicated in any theoretical sense, c) MusicBrainz and Freebase, at least (and the system I'm designing at work, I think), seem to be technically sufficient to represented the data correctly. If only we had a better plan.  
 

DiscO  

So here's my attempt at a better plan. I call it DiscO, for Discographic Ontology; that is, it's a scheme for structuring discographies. It is not an attempt at an abstract physics of human air-vibrating creativity, it is just an outline of a way to write down the information about bands, the music they've made, and how that music was released. It's intended to be simple enough that you can imagine people actually filling in the data, but expressive enough that the data can support interesting queries. And it's specifically intended to model nuance abstractly, so that it can accommodate some kinds of new needs without perpetually having to itself expand.  
 

There are four basic types:  

Artist - The Beatles, Big Country, Megadeth, Frank Zappa, whatever...  

Release - an individual album, single, compilation, whatever; Rubber Soul, Past Masters [Remastered], "Day Tripper"/"We Can Work It Out"...  

Track - an individual version of a song; "Day Tripper", "Day Tripper [mono]", "Day Tripper (performed live on pan flute and triangle by Zamfir and Elmo)", etc.  

Sequence - any collection of releases; Original Albums, Japanese Cassette Singles, 2009 Remasters, etc.  
 

These are related to each other like this:  

Artists mostly have Sequences. Sequences can be anything, but many artists would have some standard ones: Original Albums, Singles, Compilations, Remastered Albums, Current Catalog.  

Sequences have Releases (and Artists).  

Releases have Dates, Labels and Tracks. A Release may have an Artist directly, but more often would have one indirectly via a Sequence.  

Releases may be related to each other via Alternate Version/Original Version links. Thus Past Masters, Vol. 1 & 2 and Past Masters [Remastered] are both Releases, but Past Masters [Remastered] has an Original Version link to Past Masters, Vol. 1 & 2, and Past Masters, Vol. 1 & 2 has an Alternate Version link to Past Masters [Remastered].  

Tracks have Durations. A Track may have an Artist directly (so individual tracks on multi-artist compilations can be attributed correctly), but more often would have one indirectly via Release (which itself may have one indirectly via Sequence).  

Tracks may also be related to each other via Alternate Version/Original Version links. "Day Tripper" and "Day Tripper [mono]" are both Tracks, but "Day Tripper" has an Alternate Version link to "Day Tripper [mono]", and "Day Tripper [mono]" has an Original Version link to "Day Tripper". (We can get into geek arguments about which versions are the same and which are derivations (of which!), if we want, but whatever we decide, we can model.)  

Restated in schema-ish form, that's:  

Artist
- Sequence
- Release
- Track  

Sequence
- Artist
- Release  

Release
- Sequence
- Artist
- Date
- Label
- Track
- Original Version
- Alternate Version  

Track
- Artist
- Duration
- Original Version
- Alternate Version  

I think that's basically enough. What it gives up in expressiveness, it gains in usability. Our Beatles data can now, I think, be modeled both tractably and informatively. We can hook up all the versions of albums and versions of songs. We can create whatever sequences we need, and since the sequences themselves are just data, it's fine to have "Canadian Singles" for the Beatles and "Fanzine Flexis" for The Bedsitters without implying that either band should also have the other.  

And using Thread, the query-language I will (before long, hopefully) be attempting to spread through the universe, we can start to ask our questions in a way the computer can answer:  

Track:=Day Tripper.Release
 

This is our naive query. It gets all the releases that have any track called exactly "Day Tripper". Good for assuring us there's some data in the bucket, but not much help in answering our question.  

Track:=Day Tripper.Release:(.Artist:=The Beatles)
 

That limits our results to albums by the Beatles, but there are still too many. With our fully-interlinked data-model, though, we can now actually ask something that is much closer to what we mean:  

Artist:=The Beatles.Sequence:=Current Catalog.Release:(.Track:=Day Tripper)
 

That is, find the artist The Beatles, get their Current Catalog sequence, get that sequence's releases, and filter those releases down to the ones that contain a track called exactly "Day Tripper". This is progress.  

But "called exactly 'Day Tripper'" will exclude "Day Tripper [mono]", which isn't what we want to do. We're trying to ask a musical question about a song, not a typographical question about a title. But this, too, we have the powers to cope with:  

Track|Day Trippers=(...__,Original Version:=Day Tripper)  

Artist:=The Beatles.Sequence:=Current Catalog.Release:(.Track.Day Trippers)|
DT Versions=(.Track:Day Trippers=?),
Other Tracks=(.Track:Day Trippers=_._Count)
 

This time we first define a new inferred relationship on Track called "Day Trippers", which gets the Track, all its Original Versions, all their Original Versions (recursively), and then filters this set of tracks down to just the ones called "Day Tripper".  

Then we get the Beatles' current catalog releases again, but this time instead of checking each release for a track named "Day Tripper", we use our Day Trippers relationship to check for a track that is, or is derived from, "Day Tripper". And then, for each of the releases that have one, we infer two new relationships: "DT Versions" tells us which track(s) on this release are versions of "Day Tripper", and "Other Tracks" counts the tracks on this release that are not derivations of "Day Tripper".  

I.e.:  

# Release DT Versions Other Tracks
1Past Masters [Remastered] Day Tripper 32
2Mono Box Set Day Tripper [mono] 212
3Stereo Box Set Day Tripper 238
 

So now we know our choices. It took us so long to find out, but we found out.  
 
 
 

Tantalizing Postscript: But now that we have these three options before us, how do their contents overlap or differ, track by track?! We could bring up three different windows and squint at them. Or we could ask the computer:  

Artist:=The Beatles.Sequence:=Current Catalog.Release:(.Track.Day Trippers)
/(.Track...__,Original Version:##1)/=nodes
 

Aaah. I see now. (How nice it will be when I'm allowed to show you...)
Recognition: "Pitatoes!"
Speculation: "Lyra have pitatoes?"
Direction: "Lyra have pitatoes on OWN PLATE!"
Clarification: "Lyra have pitatoes on own LITTLE plate RIGHT NOW!"
Description: "Taste like bannas!"
Extrapolation: "Daddy's pitatoes taste like bannas?"
Invention: "Lyra make more pitatoes in da PLATE!" (Stirs remaining sweet potatoes vigorously.)
Context: "Chicken TOO BIG! CUT IT!" (Small piece of chicken is bisected; both halves are immediately crammed in mouth at once.)
Misdirection: "Affer dinner, have yogit pop OUTSIDE?"
Capitalization: "ALL DONE WIF DINNER!! Have yogit pop OUTSIDE!"
Since I've done this before for a couple other search engines, here's a side-by-side Bing vs Google comparator.  

I haven't tried enough things to have a very considered opinion, but here are the first few tests that went into my preliminary unimpressedness with Bing:  

kate bush covers
president twice
president obama's birthday
thread query language
where the rose is sown
capital of estonia
boston to asheville
36 jfk street
austin population
primitons reissue
ellen barkin buckaroo banzai picture  

[11 June note: this post itself is now (temporarily?) towards the top of Google's results for some of these terms! But not Bing's. Good demonstration of the observer effect, at least.]
In his post explaining his departure from Google, Douglas Bowman says "Yes, it's true that a team at Google couldn't decide between two blues, so they're testing 41 shades between each blue to see which one performs better. I had a recent debate over whether a border should be 3, 4 or 5 pixels wide, and was asked to prove my case. I can't operate in an environment like that."  

He's not really trying to have an opinion-changing last word in an argument against data-driven product decisions, but if he were, this is not how to do it. If you believe that the proper width of a border can be tested, then Bowman's refusal to subject his intuitions to quantitative confirmation just sounds like petulant prima-donna nonsense. If you can test 41 shades of blue, this line of reasoning goes, you don't need to guess, so a guessing specialist is an annoying waste of everybody's time.  

The great advantage of testing and data, of course, is that you get precise, decisive answers you can act on. Shade 31, with 3.7%, trouncing runner-up shade 14 with only 3.4%! Apply shade 31, declare progress.  

But the great disadvantage of testing and data is that you get precise, decisive answers you can and will act on, but you almost never know what question you really asked. Sure, the people who saw shade 31 did some measurable thing at some measurable rate. But why? Is it shade 31? Or is it the constrast between shade 31 and the old shade? Or is it the interplay between shade 31 and some other thing you aren't thinking about, or possibly don't even control? Are you going to run your test again in a month to see if the results have changed? Did you cross-correlate them with HTTP_REFERER and index the colors on the pages people came from? What about all the combinations of these 41 shades and 41 backgrounds and 8 fonts and 3 border widths (12 if you vary each side of the box separately!) and 41 line-heights and 19 title-bar wordings and the color of the tie Jon Stewart was wearing the night before? Which things matter? How do you know?  

You don't. And if you need to add some new element, tomorrow, you don't know which of the tests you've already run it invalidates. Are you going to rerun all of them, permuted 41 more new ways?  

No. You are going to sheepishly post a job opening for a new guessing specialist. Bowman already had his last word. It was "Goodbye".
Between this (first song on the album) and this (listen to the lyrics) (the latter by this band), that's now two bands I once wrote about (and who know that I wrote about them) who have later used my column's title in songs! Neither seems to be about me or the column, but still...
An adequate computer language allows humans to communicate with machines about machine concerns.  

A good computer language also facilitates communication between humans about machine concerns.  

A great language allows machines to participate in conversations between humans about human concerns.  
 

There are not very many of this last sort. As I've mentioned before, I'm trying to write one. I've been calling it a query language, but I've started to think I shouldn't. It's a language for talking about data-relationships, where most other things called "query languages" are for excerpting data, and the two are different qualitative goals even when the individual tasks end up being logistically similar. I'm trying to do for data-relationships what the system for symbolic algebra did for numbers. Not what algebra did for numbers, thankfully, just what we accomplished by making up a written syntax for expressing algebra compactly and precisely.  
 

So here's just one real-world example from yesterday. We were talking, elsewhere, about how you calculate overall ratings for bands in a large reviews database. The simplest thing is just to average all their ratings. In Thread, my data-relationship language, this is:  

Artist|(.Album.Rating._Average)
 

I.e.: For each artist, get their albums, then get those albums' ratings, then average all the ratings. But this is maybe not the best statistic, as it weights albums proportionally to the number of reviews. Maybe we want to average the ratings for each album, and then average the album-averages to get the artist average. That's a hard sentence for a person to read, and the computer can't read it at all. But in Thread it's just:  

Artist|(.Album.(.Rating._Average)._Average)
 

Run this, though, and you see that the top of the list is dominated by bands with very small numbers of very high ratings. Not really what we're trying to find out. So let's include only bands with at least 25 ratings:  

Artist:(.Album.Rating::#25)|(.Album.(.Rating._Average)._Average)
 

This is better, but maybe not as much better as you'd think. It turns out that there are a number of bands for which a small number of people have written a large number of reviews. Maybe what we really want is to average the ratings for each user, not for each album. That way one person giving the same high rating to 8 different albums counts as 1, not 8. And we'll only consider artists with ratings from 25 different users, not just 25 ratings total. This is:  

Artist:|(.Album.Rating/User::#25.(.group._Average)._Average)
 

Better, but it's still pretty easy to game this by creating new accounts and filing one very high rating from each of them. We can mitigate that, though, by trusting only ratings from users who have rated, say, at least 5 different albums, from at least 3 different artists. That's:  

Album|Trusted Rating=(.Rating:(.User:(.Rating.Album::#5.Artist::#3)))  

Artist|(.Album.Trusted Rating/User::#25.(.group._Average)._Average)
 

Better again. But there are still a few pretty obscure things at the top of the list. This doesn't prove that the results are flawed, of course, but scrutinizing them, and thinking about the sample-size effects of rating variation at this scale, reveals that the highest and lowest ratings are having pretty dramatic effects. Perhaps it would be smart to toss out the top and bottom 10% of the per-reviewer averages, averaging only the middle 10%. This keeps one perspective-challenged fan or one vengeful ex-bassist from single-handedly jumping the ratings up or down. Thus:  

Album|Trusted Rating=(.Rating:(.User:(.Rating.Album::#5.Artist::#3)))  

Artist|(.Album.Trusted Rating/User::#25.(.group._Average)#._Trim 10%._Average)
 

The result of this, in fact, is this leaderboard. By these rules Immolation is currently the top-ranked band in the Encyclopaedia Metallum.  
 

The English version of this final formulation is "bands with 25+ reviewers of their full-length albums, counting only reviewers who have filed at least 5 reviews and covered at least 3 bands; scored by averaging the ratings from each reviewer, dropping the top and bottom 10% of these reviewer-averages, and then averaging the remainder". This is a long sentence for people, and a useless sentence for machines, and as long as this is our canonical format, we will be at considerable risk for error every time we retranslate into a computer language. Put this in SPARQL or SQL or MQL, though, and it would be essentially inaccessible to people. So you chose between knowing what you want and not necessarily getting it, or knowing what you're getting but not whether it's what you want.  

I think we have to do better. The human stakes for data-comprehension are approaching critical levels, and our tools have not kept up. Worse, the shiny new tools in the big labs are not ready yet and not even that great.  

So Thread is my own personal attempt at doing better. Could it be the language we could actually share, humans and computers, to talk about data? I can't prove it is yet, and the project in which it's embedded is still working towards its public debut, so you can't make up your own mind yet, either. But for the past couple years I've been using it to talk to computers, and to myself, and even to a few coworkers, and the experience at least gives me hope. I know it's powerful, and I know it's compact.  

Like any language, of course, we'd have to learn it. I make no claims of it being "intuitive", whatever meaning that term might have for a symbolic-reasoning language, nor do I claim it's trivially implemented at scale. It's cryptic in its own particular way, and poses its own technical challenges. But I'm not trying to minimize anybody's absolute difficulty, I'm trying to maximize the ratio of power to difficulty. If, reading those examples above, without a formal tutorial or even an actual diagram of the data model in question, you have at least a sense of what might be going on, then it's at least possible I'm getting somewhere.  
 

[Note from a few days later: in re-reading these queries I actually noticed a methodological error! The first time I did this, I neglected to sort the ratings before trimming the first and last few. That is, I did this:  

Album|Trusted Rating=(.Rating:(.User:(.Rating.Album::#5.Artist::#3)))  

Artist|(.Album.Trusted Rating/User::#25.(.group._Average)._Trim 10%._Average)
 

where I should have done this:  

Album|Trusted Rating=(.Rating:(.User:(.Rating.Album::#5.Artist::#3)))  

Artist|(.Album.Trusted Rating/User::#25.(.group._Average)#._Trim 10%._Average)
 

The operative difference is the "#" for sorting right before "._Trim 10%" in the second query, which is what makes the trim function take off the highest and lowest ratings, rather than just the first and last.  

But even this error is kind of my point. The language is a tool for me to talk to myself over time.]
1. The tendency to forget completely about anything for which someone else is vaguely expected to take the next step.  

2. A reluctance to accept that quantifying one's nostalgia does not mitigate its mortality.  

3. A vigilant willingness to challenge abstrusely tangential orthodoxies.  

4. A failure, when not concentrating, to properly aspirate the letter H in the words "humor" and "human".  

5. A fear of widths.  

6. The maintenance of short but meticulous lists of inconclusive evidence for undeniable truths.  

7. Always, or almost, allowing the silent moments at the ends of experiences to complete without crossfade.
Site contents published by glenn mcdonald under a Creative Commons BY/NC/ND License except where otherwise noted.