Songlinks › How it decides

How Songlinks decides whether two songwriter records are one person

Two registrations go in. First a gate: do they share any catalogue that counts? Then a chain of questions about the evidence, each answer a push toward “same” or “different”. Many such chains put their pushes on a pair of scales, and the needle settles in one of three zones. Then many pairs at once.

Two records, a gate, a chain of questions, a pair of scales, a needle. The animation needs a browser with WebGL; the words below say the same thing.
held for review

What you just watched

The gate

Songlinks compares records two at a time, but only when they share a publisher, co-writer or administrator that counts (not a society, a major or an aggregator). Without that, no score is computed and nothing is written: a name alone is never evidence. That is most pairs, and it is on purpose: Songlinks abstains rather than guess.

The board is a chain of questions

Each peg asks one thing about the pair: do they share a publisher, do they share co-writers, is the spelling rare, do both carry an IPI name number. The ball’s route is the pair’s answers, and the slot it lands in is a push toward “same” or “different”. That push is the chain’s whole say.

The scales are all the other chains

There are many chains of questions, and every one of them puts a weight on the “same” pan or the “different” pan. The needle is their sum. No single question, and no name match on its own, can move it far.

The zones are the rules

In the green the two records are joined. In the amber nothing is written: a person opens both records and decides. In the grey they stay apart. The bar for joining is high on purpose, because a wrong join pays the wrong person.

The four pairs are made up, names included, to show the four outcomes. The questions are the kinds of evidence Songlinks weighs, in plain words; the pushes are sizes, not scores. What the model is, which questions it asks in what order and where its lines are drawn is not published, as About explains; the FAQ says what kind of model it is.

Decision trees, from the top

A tree is a chain of questions

Think of the game of twenty questions. Each question has a yes or a no, each answer chooses the next question, and after a few of them you have narrowed a whole world of possibilities to one guess. A decision tree is that, written down: a first question at the top, a question for each answer below it, and at the bottom, where the questions run out, a leaf that holds the guess. Walking a tree costs nothing: a handful of comparisons, each one readable. The chains in the animation are decision trees, with a peg for every question and a slot for every leaf.

How a tree is learned

Nobody writes the questions by hand. A tree is grown from examples whose answer is already known: here, pairs of records that a person or an identifier has settled as the same human or not. Growing starts at the top with every example in one heap and tries each question it could ask, keeping the one that splits the heap most cleanly into “same” and “different”. Then it does the same for each half, and again for each quarter, until the heaps are too small or too pure to be worth splitting. What is left in each leaf is the answer the tree gives to anything that lands there. The questions, their order, and the lines they draw are all learned from the examples, which is why they are the part of a system that is worth keeping to yourself.

Why one tree is not enough

A single tree is easy to read and easy to fool. It is brittle: change a few examples and a question near the top can flip, and everything below it changes with it. It also memorises: given enough questions it can carve the examples it grew from into perfect leaves and have learned nothing that carries to a pair it has never seen. The usual cure is many trees, each kept short and each given only a small say, with the final answer their sum. One short chain can be wrong about a pair; it is much harder for most of them to be wrong the same way. That is what the scales are doing: every chain puts its small weight on a pan, and the needle reads the total.

What the needle is, and what it is not

The total is a score. Scores from a system like this are usually lined up against the past so that a position on the dial means “of the pairs that landed here before, this share turned out to be the same person”; that is what the dial is drawn to suggest. The zones are not part of the tree at all. They are rules laid on top of the score: where joining is allowed, where a person must look, where nothing happens. Those rules are a choice about the cost of a mistake, not a thing the examples can teach, and the amber band exists because a wrong join costs more than a missed one.

What a tree cannot do

A tree only knows the evidence it is handed. It does not read a name the way a person does, does not know who played on what, and has no sense of what is plausible beyond the examples it grew from. In particular, evidence that is missing is not evidence of difference: two records that share no publisher may be the same writer who was never registered through one. That is why the gate comes before the trees. A pair with nothing to weigh is not judged at all, instead of being judged on the absence. And a tree is only as good as its examples: if the settled pairs are mostly from one part of the world or one era, the questions it learns will be good at that part and poor elsewhere, without ever saying so.

Going further

Two places to learn this properly, both free to start and neither ours: