Skip to content

Distance metrics such as cosine similarity don't work #33

Open
@TheColorman

Description

@TheColorman

When using a metric such as Cosine similarity, the returned vector is almost never the one with the smallest distance value.
Example:

Dataset
var points = [
{ 'x': 61, 'y': 54, 'id': 'A' },
{ 'x': 7, 'y': 54, 'id': 'B' },
{ 'x': 62, 'y': 58, 'id': 'C' },
{ 'x': 75, 'y': 48, 'id': 'D' },
{ 'x': 76, 'y': 52, 'id': 'E' },
{ 'x': 41, 'y': 16, 'id': 'F' },
{ 'x': 35, 'y': 35, 'id': 'G' },
{ 'x': 34, 'y': 63, 'id': 'H' },
{ 'x': 22, 'y': 32, 'id': 'I' },
{ 'x': 0, 'y': 53, 'id': 'J' },
{ 'x': 13, 'y': 91, 'id': 'K' },
{ 'x': 87, 'y': 24, 'id': 'L' },
{ 'x': 63, 'y': 2, 'id': 'M' },
{ 'x': 10, 'y': 76, 'id': 'N' },
{ 'x': 85, 'y': 98, 'id': 'O' },
{ 'x': 94, 'y': 90, 'id': 'P' },
{ 'x': 52, 'y': 47, 'id': 'Q' },
{ 'x': 100, 'y': 38, 'id': 'R' },
{ 'x': 95, 'y': 47, 'id': 'S' },
{ 'x': 62, 'y': 15, 'id': 'T' },
{ 'x': 93, 'y': 66, 'id': 'U' },
{ 'x': 46, 'y': 54, 'id': 'V' },
{ 'x': 85, 'y': 99, 'id': 'W' },
{ 'x': 32, 'y': 53, 'id': 'X' },
{ 'x': 11, 'y': 37, 'id': 'Y' },
{ 'x': 0, 'y': 54, 'id': 'Z' },
{ 'x': 90, 'y': 100, 'id': 's' },
{ 'x': 84, 'y': 58, 'id': 't' },
{ 'x': 97, 'y': 35, 'id': 'u' },
{ 'x': 24, 'y': 34, 'id': 'v' },
{ 'x': 67, 'y': 70, 'id': 'w' },
{ 'x': 16, 'y': 7, 'id': 'x' },
{ 'x': 27, 'y': 73, 'id': 'a' },
{ 'x': 0, 'y': 35, 'id': 'b' },
{ 'x': 97, 'y': 47, 'id': 'c' },
{ 'x': 44, 'y': 56, 'id': 'd' },
{ 'x': 23, 'y': 11, 'id': 'e' },
{ 'x': 3, 'y': 80, 'id': 'f' },
{ 'x': 87, 'y': 27, 'id': 'g' },
{ 'x': 42, 'y': 29, 'id': 'h' },
{ 'x': 77, 'y': 72, 'id': 'i' },
{ 'x': 40, 'y': 10, 'id': 'j' },
{ 'x': 86, 'y': 91, 'id': 'k' },
{ 'x': 43, 'y': 23, 'id': 'l' },
{ 'x': 55, 'y': 13, 'id': 'm' },
{ 'x': 88, 'y': 14, 'id': 'n' },
{ 'x': 67, 'y': 22, 'id': 'o' },
{ 'x': 88, 'y': 91, 'id': 'p' },
{ 'x': 82, 'y': 33, 'id': 'q' },
{ 'x': 97, 'y': 29, 'id': 'r' }
]
Cosine similarity (note the negative sign at the end of the function, as it would normally increase with similar vectors):
    const cosineSimilarity = (vec1, vec2) => {
        let dotProduct = 0;
        let magnitude1 = 0;
        let magnitude2 = 0;

        for (const dim of Object.keys(vec1)) {
            if (typeof (vec1[dim]) != 'number') continue;

            dotProduct += vec1[dim] * vec2[dim];
            magnitude1 += Math.pow(vec1[dim], 2);
            magnitude2 += Math.pow(vec2[dim], 2);
        }

        return -dotProduct / (Math.sqrt(magnitude1) * Math.sqrt(magnitude2));
    };

Inefficient function the get the closest vectors from an array.

    const closestVectors = (vec, vectors) => {
        const distances = vectors.map(v => cosineSimilarity(vec, v));
        const min = Math.min(...distances);
        // Check if multiple with same distance
        const minIndices = distances.reduce((a, e, i) => {
            if (e === min) a.push(i);
            return a;
        }, []);
        // Return list of closest vectors
        return { vectors: minIndices.map(i => vectors[i]), distances: minIndices.map(i => distances[i]) };
    };

Testing

    const test = { x: 200, y: 123, label: "TEST" }
    var tree = new kdTree(points, cosineSimilarity, ['x', 'y']);

    const nearest = tree.nearest(test, 1);
    const nearest2 = closestVectors(test, points);

    console.log("K-D Tree")
    console.log(nearest)
    console.log("Custom")
    console.log(nearest2)

  // > K-D Tree
  // > [ [ { x: 76, y: 52, id: 'E' }, -0.998815642613221 ] ]
  // > Custom
  // > {
  //     vectors: [ { x: 75, y: 48, id: 'D' } ],
  //     distances: [ -0.999839132267399 ]
  //   }

Clearly the tree returns a vector with a slightly larger distance (0.998 vs 0.999).

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions