apr144-hclust
v0.0.1
Published
Agglomerative hierarchical clustering in JavaScript
Downloads
7
Maintainers
Readme
hclust
Agglomerative hierarchical clustering in JavaScript
Inspired by the MIT-licensed hcluster.js by @ChrisPolis. See the comparison of the two below.
Usage
Browser
<script src="hclust.min.js"></script>
<script>
hclust.clusterData(...);
hclust.euclideanDistance(...);
hclust.avgDistance(...);
</script>
Node
npm install @greenelab/hclust
or
yarn add @greenelab/hclust
then
import { clusterData } from 'hclust';
import { euclideanDistance } from 'hclust';
import { avgDistance } from 'hclust';
clusterData({ data, key, distance, linkage, onProgress })
Parameters
data
The data you want to cluster, in the format:
[
...
[ ... 1, 2, 3 ...],
[ ... 1, 2, 3 ...],
[ ... 1, 2, 3 ...],
...
]
or if key
parameter is specified:
[
...
{ someKey: [ ... 1, 2, 3 ...] },
{ someKey: [ ... 1, 2, 3 ...] },
{ someKey: [ ... 1, 2, 3 ...] },
...
]
The entries in the outer array can be considered the rows
and the entries within each row
array can be considered the cols
.
Each row
should have the same number of cols
.
Default value: []
key
A string
key to specify which values to extract from the data
array.
If omitted, data
is assumed to be an array of arrays.
If specified, data
is assumed to be array of objects, each with a key that contains the values for that row
.
Default value: ''
distance
A function to calculate the distance between two equal-dimension vectors, used in calculating the distance matrix, in the format:
function (arrayA, arrayB) { return someNumber; }
The function receives two equal-length arrays of numbers (ints or floats) and should return a number (int or float).
Default value: euclideanDistance
from this hclust
package
linkage
A function to calculate the distance between pairs of clusters based on a distance matrix, used in determining linkage criterion, in the format:
function (arrayA, arrayB, distanceMatrix) { return someNumber; }
The function receives two sets of indexes and the distance matrix computed between each datum and every other datum. The function should return a number (int or float)
Default value: averageDistance
from this hclust
package
onProgress
A function that is called several times throughout clustering, and is provided the current progress through the clustering, in the format:
function (progress) { }
The function receives the percent progress between 0
and 1
.
Default value: an internal function that console.log
's the progress
Note: postMessage
is called in the same places as onProgress
, if the script is running as a web worker.
Returns
const { clusters, distances, order, clustersGivenK } = clusterData(...);
clusters
The resulting cluster tree, in the format:
{
indexes: [ ... Number, Number, Number ... ],
height: Number,
children: [ ... {}, {}, {} ... ]
}
distances
The computed distance matrix, in the format:
[
...
[ ... Number, Number, Number ...],
[ ... Number, Number, Number ...],
[ ... Number, Number, Number ...]
...
]
order
The new order of the data, in terms of original data array indexes, in the format:
[ ... Number, Number, Number ... ]
Equivalent to clusters.indexes
and clustersGivenK[1]
.
clustersGivenK
A list of tree slices in terms of original data array indexes, where index = K, in the format:
[
[], // K = 0
[ [] ], // K = 1
[ [], [] ], // K = 2
[ [], [], [] ], // K = 3
[ [], [], [], [] ], // K = 4
[ [], [], [], [], [] ] // K = 5
...
]
euclideanDistance(arrayA, arrayB)
Calculates the euclidean distance between two equal-dimension vectors.
avgDistance(arrayA, arrayB, distanceMatrix)
Calculates the average distance between pairs of clusters based on a distance matrix.
Comparison with hcluster.js
- This package does not duplicate items from the original dataset in the results. Results are given in terms of indexes, either with respect to the original dataset or the distance matrix.
- This package uses more modern JavaScript syntaxes and practices to make the code cleaner and simpler.
- This package provides an
onProgress
callback and callspostMessage
for use in web workers. Because clustering can take a long time with large data sets, you may want to run it as a web worker so the browser doesn't freeze for a long time, and you may need a callback so you can give users visual feedback on its progress. - This package makes some performance optimizations, such as removing unnecessary loops through big sets.
It has been tested on various OS's (Windows, Mac, Linux, iOS, Android), devices (desktop, laptop, mobile), browsers (Chrome, Firefox, Safari), contexts (main thread, web worker), and hosting locations (local, online).
The results vary widely, and are likely sensitive to the specifics of hardware, cpu cores, browser implementation, etc.
But in general, this package is more performant than
hcluster.js
, to varying degrees, and is always at least as performant on average. Chrome seems to see the most performance gains (up to 10x, when the row number is high), while Firefox seems to see no gains. - This package does not touch the input data object, whereas the
hcluster.js
package does. D3 often expects you to mutate data objects directly, which is now typically considered bad practice in JavaScript. Instead, this package returns the useful data from the clustering algorithm (including the distance matrix), and allows you to mutate or not mutate the data object depending on your needs. - This package leaves out the
minDistance
ormaxDistance
functions that are built intohcluster.js
, because -- per this reference -- they are not as effective asaverageDistance
.
Making changes to the library
- Install Node
- Install Yarn
- Clone this repo and navigate to it in your command terminal
- Run
yarn install
to install this package's dependencies - Make desired changes to
./src/hclust.js
- Run
yarn test
to automatically rebuild the library and run test suite - Run
yarn build
to just rebuild the library, and output the compiled contents to./build/hclust.min.js
- Commit changes to repo if necessary. Make sure to run the build command before committing; it won't happen automatically.
Similar libraries
cmpolis/hcluster.js harthur/clustering mljs/hclust math-utils/hierarchical-clustering
Further reading
The AGNES (AGglomerative NESting) method; continuously merge nodes that have the least dissimilarity.