1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309
use crate::dbscan::{DbscanParams, DbscanValidParams};
use linfa_nn::{
distance::{Distance, L2Dist},
CommonNearestNeighbour, NearestNeighbour, NearestNeighbourIndex,
};
use ndarray::{Array1, ArrayBase, Data, Ix2};
use std::collections::VecDeque;
use linfa::Float;
use linfa::{traits::Transformer, DatasetBase};
#[derive(Clone, Debug, PartialEq, Eq)]
/// DBSCAN (Density-based Spatial Clustering of Applications with Noise)
/// clusters together points which are close together with enough neighbors
/// labelled points which are sparsely neighbored as noise. As points may be
/// part of a cluster or noise the predict method returns
/// `Array1<Option<usize>>`
///
/// As it groups together points in dense regions the number of clusters is
/// determined by the dataset and distance tolerance not the user.
///
/// We provide an implemention of the standard O(N^2) query-based algorithm
/// of which more details can be found in the next section or
/// [here](https://en.wikipedia.org/wiki/DBSCAN).
///
/// The standard DBSCAN algorithm isn't iterative and therefore there's
/// no fit method provided only predict.
///
/// ## The algorithm
///
/// The algorithm iterates over each point in the dataset and for every point
/// not yet assigned to a cluster:
/// - Find all points within the neighborhood of size `tolerance`
/// - If the number of points in the neighborhood is below a minimum size label
/// as noise
/// - Otherwise label the point with the cluster ID and repeat with each of the
/// neighbours
///
/// ## Tutorial
///
/// Let's do a walkthrough of an example running DBSCAN on some data.
///
/// ```rust
/// use linfa::traits::*;
/// use linfa_clustering::{DbscanParams, Dbscan};
/// use linfa_datasets::generate;
/// use ndarray::{Axis, array, s};
/// use ndarray_rand::rand::SeedableRng;
/// use rand_xoshiro::Xoshiro256Plus;
/// use approx::assert_abs_diff_eq;
///
/// // Our random number generator, seeded for reproducibility
/// let seed = 42;
/// let mut rng = Xoshiro256Plus::seed_from_u64(seed);
///
/// // `expected_centroids` has shape `(n_centroids, n_features)`
/// // i.e. three points in the 2-dimensional plane
/// let expected_centroids = array![[0., 1.], [-10., 20.], [-1., 10.]];
/// // Let's generate a synthetic dataset: three blobs of observations
/// // (100 points each) centered around our `expected_centroids`
/// let observations = generate::blobs(100, &expected_centroids, &mut rng);
///
/// // Let's configure and run our DBSCAN algorithm
/// // We use the builder pattern to specify the hyperparameters
/// // `min_points` is the only mandatory parameter.
/// // If you don't specify the others (e.g. `tolerance`)
/// // default values will be used.
/// let min_points = 3;
/// let clusters = Dbscan::params(min_points)
/// .tolerance(1e-2)
/// .transform(&observations)
/// .unwrap();
/// // Points are `None` if noise `Some(id)` if belonging to a cluster.
/// ```
///
pub struct Dbscan;
impl Dbscan {
/// Configures the hyperparameters with the minimum number of points required to form a cluster
///
/// Defaults are provided if the optional parameters are not specified:
/// * `tolerance = 1e-4`
/// * `dist_fn = L2Dist` (Euclidean distance)
/// * `nn_algo = KdTree`
pub fn params<F: Float>(min_points: usize) -> DbscanParams<F, L2Dist, CommonNearestNeighbour> {
Self::params_with(min_points, L2Dist, CommonNearestNeighbour::KdTree)
}
/// Configures the hyperparameters with the minimum number of points, a custom distance metric,
/// and a custom nearest neighbour algorithm
pub fn params_with<F: Float, D: Distance<F>, N: NearestNeighbour>(
min_points: usize,
dist_fn: D,
nn_algo: N,
) -> DbscanParams<F, D, N> {
DbscanParams::new(min_points, dist_fn, nn_algo)
}
}
impl<F: Float, D: Data<Elem = F>, DF: Distance<F>, N: NearestNeighbour>
Transformer<&ArrayBase<D, Ix2>, Array1<Option<usize>>> for DbscanValidParams<F, DF, N>
{
fn transform(&self, observations: &ArrayBase<D, Ix2>) -> Array1<Option<usize>> {
let mut cluster_memberships = Array1::from_elem(observations.nrows(), None);
let mut current_cluster_id = 0;
// Tracks whether a value is in the search queue to prevent duplicates
let mut search_found = vec![false; observations.nrows()];
let mut search_queue = VecDeque::with_capacity(observations.nrows());
// Construct NN index
let nn = match self.nn_algo.from_batch(observations, self.dist_fn.clone()) {
Ok(nn) => nn,
Err(linfa_nn::BuildError::ZeroDimension) => {
return Array1::from_elem(observations.nrows(), None)
}
Err(e) => panic!("Unexpected nearest neighbour error: {}", e),
};
for i in 0..observations.nrows() {
if cluster_memberships[i].is_some() {
continue;
}
let (neighbor_count, neighbors) =
self.find_neighbors(&*nn, i, observations, self.tolerance, &cluster_memberships);
if neighbor_count < self.min_points {
continue;
}
neighbors.iter().for_each(|&n| search_found[n] = true);
search_queue.extend(neighbors.into_iter());
// Now go over the neighbours adding them to the cluster
cluster_memberships[i] = Some(current_cluster_id);
while let Some(candidate_idx) = search_queue.pop_front() {
search_found[candidate_idx] = false;
let (neighbor_count, neighbors) = self.find_neighbors(
&*nn,
candidate_idx,
observations,
self.tolerance,
&cluster_memberships,
);
// Make the candidate a part of the cluster even if it's not a core point
cluster_memberships[candidate_idx] = Some(current_cluster_id);
if neighbor_count >= self.min_points {
for n in neighbors.into_iter() {
if !search_found[n] {
search_queue.push_back(n);
search_found[n] = true;
}
}
}
}
current_cluster_id += 1;
}
cluster_memberships
}
}
impl<F: Float, D: Data<Elem = F>, DF: Distance<F>, N: NearestNeighbour, T>
Transformer<
DatasetBase<ArrayBase<D, Ix2>, T>,
DatasetBase<ArrayBase<D, Ix2>, Array1<Option<usize>>>,
> for DbscanValidParams<F, DF, N>
{
fn transform(
&self,
dataset: DatasetBase<ArrayBase<D, Ix2>, T>,
) -> DatasetBase<ArrayBase<D, Ix2>, Array1<Option<usize>>> {
let predicted = self.transform(dataset.records());
dataset.with_targets(predicted)
}
}
impl<F: Float, D: Distance<F>, N: NearestNeighbour> DbscanValidParams<F, D, N> {
fn find_neighbors(
&self,
nn: &dyn NearestNeighbourIndex<F>,
idx: usize,
observations: &ArrayBase<impl Data<Elem = F>, Ix2>,
eps: F,
clusters: &Array1<Option<usize>>,
) -> (usize, Vec<usize>) {
let candidate = observations.row(idx);
let mut res = Vec::with_capacity(self.min_points);
let mut count = 0;
// Unwrap here is fine because we don't expect any dimension mismatch when calling
// within_range with points from the observations
for (_, i) in nn.within_range(candidate.view(), eps).unwrap().into_iter() {
count += 1;
if clusters[i].is_none() && i != idx {
res.push(i);
}
}
(count, res)
}
}
#[cfg(test)]
mod tests {
use super::*;
use linfa::ParamGuard;
use linfa_nn::{distance::L1Dist, BallTree};
use ndarray::{arr1, arr2, s, Array2};
#[test]
fn nested_clusters() {
// Create a circuit of points and then a cluster in the centre
// and ensure they are identified as two separate clusters
let mut data: Array2<f64> = Array2::zeros((50, 2));
let rising = Array1::linspace(0.0, 8.0, 10);
data.column_mut(0).slice_mut(s![0..10]).assign(&rising);
data.column_mut(0).slice_mut(s![10..20]).assign(&rising);
data.column_mut(1).slice_mut(s![20..30]).assign(&rising);
data.column_mut(1).slice_mut(s![30..40]).assign(&rising);
data.column_mut(1).slice_mut(s![0..10]).fill(0.0);
data.column_mut(1).slice_mut(s![10..20]).fill(8.0);
data.column_mut(0).slice_mut(s![20..30]).fill(0.0);
data.column_mut(0).slice_mut(s![30..40]).fill(8.0);
data.column_mut(0).slice_mut(s![40..]).fill(5.0);
data.column_mut(1).slice_mut(s![40..]).fill(5.0);
let labels = Dbscan::params(2)
.tolerance(1.0)
.check()
.unwrap()
.transform(&data);
assert!(labels.slice(s![..40]).iter().all(|x| x == &Some(0)));
assert!(labels.slice(s![40..]).iter().all(|x| x == &Some(1)));
}
#[test]
fn non_cluster_points() {
let mut data: Array2<f64> = Array2::zeros((5, 2));
data.row_mut(0).assign(&arr1(&[10.0, 10.0]));
let labels = Dbscan::params(4).check().unwrap().transform(&data);
let expected = arr1(&[None, Some(0), Some(0), Some(0), Some(0)]);
assert_eq!(labels, expected);
}
#[test]
fn border_points() {
let data: Array2<f64> = arr2(&[
// Outlier
[0.0, 2.0],
// Core point
[0.0, 0.0],
// Border points
[0.0, 1.0],
[0.0, -1.0],
[-1.0, 0.0],
[1.0, 0.0],
]);
// Run the dbscan with tolerance of 1.1, 5 min points for density
let labels = Dbscan::params(5)
.tolerance(1.1)
.check()
.unwrap()
.transform(&data);
assert_eq!(labels[0], None);
for id in labels.slice(s![1..]).iter() {
assert_eq!(id, &Some(0));
}
}
#[test]
fn l1_dist() {
let data: Array2<f64> = arr2(&[
// Outlier
[0.0, 6.0],
// Core point
[0.0, 0.0],
// Border points
[2.0, 3.0],
[1.0, -3.0],
[-4.0, 1.0],
[1.0, 1.0],
]);
// Run the L1-dist dbscan with tolerance of 5.01, 5 min points for density
let labels = Dbscan::params_with(5, L1Dist, BallTree)
.tolerance(5.01)
.check()
.unwrap()
.transform(&data);
assert_eq!(labels[0], None);
for id in labels.slice(s![1..]).iter() {
assert_eq!(id, &Some(0));
}
}
#[test]
fn dataset_too_small() {
let data: Array2<f64> = Array2::zeros((3, 2));
let labels = Dbscan::params(4).check().unwrap().transform(&data);
assert!(labels.iter().all(|x| x.is_none()));
}
}