SPB Git

Repository Public

MultivariateTests.swift 10.2 KB 197 lines swift

1import Foundation2import Testing3import AnalyticsTestSupport4@testable import Statistics56private let mvFixture = Fixture.load("multivariate", .module)78/// Renumbers labels in order of first appearance, so two labelings can be compared as partitions.9private func canonical(_ labels: [Int]) -> [Int] {10    var map = [Int: Int]()11    return labels.map { l in12        if let v = map[l] { return v }13        map[l] = map.count14        return map[l]!15    }16}1718@Suite struct MultivariatePCATests {19    let f = mvFixture["pca"]20    var x: Matrix { Matrix(f["X"].mat) }2122    /// Compares components and scores up to the sign of each component, then checks our sign convention.23    func check(_ r: PCAResult, _ ref: JSON, _ label: String, rel: Double = 1e-9) {24        let comps = ref["components"].mat, scores = ref["scores"].mat25        #expect(r.components.rows == comps.count, "\(label) component count")26        #expect(allClose(r.explainedVariance, ref["variance"].arr, rel: rel), "\(label) variance")27        #expect(allClose(r.explainedVarianceRatio, ref["ratio"].arr, rel: rel), "\(label) ratio")28        #expect(allClose(r.singularValues, ref["singular"].arr, rel: rel), "\(label) singular values")29        for (i, want) in comps.enumerated() {30            let got = r.components.row(i)31            let sign: Double = zip(got, want).reduce(0) { $0 + $1.0 * $1.1 } < 0 ? -1 : 132            #expect(allClose(got, want.map { $0 * sign }, rel: 1e-8, abs: 1e-9), "\(label) component \(i)")33            #expect(allClose(r.scores.column(i), scores.map { $0[i] * sign }, rel: 1e-7, abs: 1e-8), "\(label) scores \(i)")34            #expect(isClose(got.reduce(0) { $0 + $1 * $1 }, 1, rel: 1e-12), "\(label) unit length")35            let big = got.max { abs($0) < abs($1) }!36            #expect(big > 0, "\(label) sign convention")37        }38    }3940    @Test func covariancePCAMatchesScikitLearn() throws {41        let r = try PCA.fit(x)42        check(r, f["cov"], "cov")43        #expect(allClose(r.means, f["cov"]["mean"].arr, rel: 1e-12))44        #expect(allClose(r.explainedVariance, f["cov_matrix_eig"].arr, rel: 1e-9))45        #expect(isClose(r.cumulativeVarianceRatio.last!, 1, rel: 1e-12))46        #expect(!r.standardized && r.scales == [1, 1, 1, 1, 1])47        let r2 = try PCA.fit(x, components: 2)48        check(r2, f["cov_k2"], "cov k=2")49        #expect(r2.scores.cols == 2 && r2.loadings.rows == 5 && r2.loadings.cols == 2)50        // total variance still counts every component51        #expect(isClose(r2.totalVariance, r.explainedVariance.reduce(0, +), rel: 1e-12))52    }5354    @Test func correlationPCAMatchesScikitLearn() throws {55        let r = try PCA.fit(x, standardize: true, names: ["a", "b", "c", "d", "e"])56        check(r, f["corr"], "corr")57        #expect(allClose(r.scales, f["sd"].arr, rel: 1e-12))58        #expect(allClose(r.explainedVariance, f["corr_matrix_eig"].arr, rel: 1e-9))59        #expect(isClose(r.totalVariance, 5, rel: 1e-12))60        // loadings of a correlation PCA are the variable–component correlations61        let l = r.loadings62        for v in 0..<5 {63            for c in 0..<5 {64                #expect(isClose(l[v, c], Descriptive.pearson(x.column(v), r.scores.column(c)), rel: 1e-8, abs: 1e-9), "loading \(v),\(c)")65            }66        }67        #expect(r.table().count == 11 && r.table()[6][0] == .text("a"))68    }6970    @Test func scoresAreUncorrelatedAndTransformProjectsNewData() throws {71        let r = try PCA.fit(x)72        for i in 0..<5 {73            #expect(isClose(Descriptive.variance(r.scores.column(i)), r.explainedVariance[i], rel: 1e-9))74            for j in 0..<i { #expect(abs(Descriptive.covariance(r.scores.column(i), r.scores.column(j))) < 1e-9 * r.explainedVariance[0]) }75        }76        let proj = try r.transform(Matrix(f["new"].mat))77        let want = f["new_scores_cov"].mat78        for i in 0..<5 {79            let sign: Double = zip(r.components.row(i), f["cov"]["components"].mat[i]).reduce(0) { $0 + $1.0 * $1.1 } < 0 ? -1 : 180            #expect(allClose(proj.column(i), want.map { $0[i] * sign }, rel: 1e-8, abs: 1e-8))81        }82        #expect(allClose(try r.transform(x).data, r.scores.data, rel: 1e-10, abs: 1e-10))83        #expect(throws: StatError.self) { try r.transform(Matrix(rows: 1, cols: 3)) }84    }8586    @Test func wideDataAndErrors() throws {87        let w = Matrix(f["wide"]["X"].mat)                    // 4 observations, 7 variables → at most 3 components88        let r = try PCA.fit(w)89        #expect(r.components.rows == 3)90        #expect(allClose(r.explainedVariance, Array(f["wide"]["variance"].arr.prefix(3)), rel: 1e-9))91        #expect(throws: StatError.self) { try PCA.fit(Matrix(rows: 1, cols: 3)) }92        #expect(throws: StatError.self) { try PCA.fit(x, components: 9) }93        #expect(throws: StatError.self) { try PCA.fit(x, names: ["a"]) }94        #expect(throws: StatError.self) { try PCA.fit(Matrix([[1, 2], [1, 3], [1, 5]]), standardize: true) }95    }96}9798@Suite struct MultivariateKMeansTests {99    let f = mvFixture["kmeans"]100    var x: Matrix { Matrix(f["X"].mat) }101102    @Test func lloydFromFixedCentroidsMatchesScikitLearn() throws {103        for c in f["cases"].list {104            let k = c["k"].i105            let r = try KMeans.fit(x, k: k, initialCentroids: Matrix(c["init"].mat))106            #expect(r.converged && r.runs == 1, "k=\(k)")107            #expect(r.labels == c["labels"].ints, "k=\(k) labels")108            #expect(allClose(r.centroids.data, c["centers"].mat.flatMap { $0 }, rel: 1e-10, abs: 1e-11), "k=\(k) centres")109            #expect(isClose(r.inertia, c["inertia"].d, rel: 1e-10), "k=\(k) inertia")110            #expect(isClose(r.silhouette ?? .nan, c["silhouette"].d, rel: 1e-10), "k=\(k) silhouette")111            let s = KMeans.silhouette(x, labels: r.labels)112            #expect(allClose(s.samples, c["samples"].arr, rel: 1e-9, abs: 1e-11), "k=\(k) silhouette samples")113            #expect(r.clusterSizes.reduce(0, +) == x.rows)114        }115    }116117    @Test func seededPlusPlusFindsTheGlobalOptimumReproducibly() throws {118        let a = try KMeans.fit(x, k: 4, seed: 11, restarts: 20)119        let b = try KMeans.fit(x, k: 4, seed: 11, restarts: 20)120        #expect(a == b)121        // four well-separated groups: the best of 20 k-means++ starts reaches scikit-learn's best of 50122        #expect(isClose(a.inertia, f["best4_inertia"].d, rel: 1e-9), "\(a.inertia) vs \(f["best4_inertia"].d)")123        #expect(isClose(a.silhouette!, f["best4_silhouette"].d, rel: 1e-9))124        let three = try KMeans.fit(x, k: 3, seed: 5, restarts: 30)125        #expect(isClose(three.inertia, f["best3_inertia"].d, rel: 1e-9))126        let other = try KMeans.fit(x, k: 4, seed: 12, restarts: 1)127        #expect(other.inertia >= a.inertia - 1e-9)128        // inertia decreases with k129        let five = try KMeans.fit(x, k: 5, seed: 3, restarts: 20)130        #expect(five.inertia < a.inertia && a.inertia < three.inertia)131    }132133    @Test func plusPlusSeedsAreDataPointsAndSpreadOut() {134        var g = Xoshiro256StarStar(seed: 1)135        let c = KMeans.plusPlus(x, 4, &g)136        let rows = Set(x.rowArrays.map { $0.map(\.bitPattern) })137        for i in 0..<4 { #expect(rows.contains(c.row(i).map(\.bitPattern))) }138        #expect(Set(c.rowArrays.map { $0.map(\.bitPattern) }).count == 4)139    }140141    @Test func edgeCases() throws {142        let one = try KMeans.fit(x, k: 1)143        #expect(one.silhouette == nil && one.labels.allSatisfy { $0 == 0 })144        // inertia with one cluster = total sum of squares about the mean145        let tss = (0..<3).reduce(0.0) { $0 + Descriptive.sumSquaredDeviations(x.column($1)) }146        #expect(isClose(one.inertia, tss, rel: 1e-11))147        let small = Matrix([[0, 0], [0, 1], [10, 0], [10, 1]])148        let two = try KMeans.fit(small, k: 2, seed: 3)149        #expect(canonical(two.labels) == [0, 0, 1, 1] && two.inertia == 1)150        // an empty cluster (two identical starting centres) is re-seeded, not left empty151        let reseeded = try KMeans.fit(small, k: 2, initialCentroids: Matrix([[0, 0.5], [0, 0.5]]))152        #expect(reseeded.clusterSizes.allSatisfy { $0 > 0 })153        let all = try KMeans.fit(small, k: 4, seed: 9)154        #expect(all.inertia == 0)155        #expect(throws: StatError.self) { try KMeans.fit(small, k: 5) }156        #expect(throws: StatError.self) { try KMeans.fit(small, k: 0) }157        #expect(throws: StatError.self) { try KMeans.fit(small, k: 2, initialCentroids: Matrix(rows: 3, cols: 2)) }158        #expect(two.table().count == 8)159    }160}161162@Suite struct MultivariateHierarchicalTests {163    let f = mvFixture["hier"]164165    @Test func linkageMatchesSciPy() throws {166        let x = Matrix(f["X"].mat)167        for method in HierarchicalClustering.Linkage.allCases {168            let ref = f[method.rawValue]169            let r = try HierarchicalClustering.linkage(x, method: method)170            let z = ref["Z"].mat171            #expect(r.merges.count == z.count && r.n == 45)172            for (m, row) in zip(r.merges, z) {173                #expect(m.first == Int(row[0]) && m.second == Int(row[1]), "\(method) merge order")174                #expect(isClose(m.distance, row[2], rel: 1e-10), "\(method) distance")175                #expect(m.size == Int(row[3]))176            }177            #expect(canonical(try r.labels(clusters: 3)) == canonical(ref["labels3"].ints), "\(method) 3 clusters")178            #expect(canonical(try r.labels(clusters: 5)) == canonical(ref["labels5"].ints), "\(method) 5 clusters")179            #expect(Set(try r.labels(clusters: 45)).count == 45)180            #expect(Set(try r.labels(clusters: 1)) == [0])181        }182    }183184    @Test func handExample() throws {185        // points on a line: 0, 1, 3, 7186        let r = try HierarchicalClustering.linkage(Matrix([[0], [1], [3], [7]]), method: .single)187        #expect(r.merges.map(\.distance) == [1, 2, 4])188        #expect(r.merges[0].first == 0 && r.merges[0].second == 1 && r.merges[1].first == 2 && r.merges[1].second == 4)189        let c = try HierarchicalClustering.linkage(Matrix([[0], [1], [3], [7]]), method: .complete)190        #expect(c.merges.map(\.distance) == [1, 3, 7])191        #expect(try r.labels(clusters: 2) == [0, 0, 0, 1])192        #expect(r.table().count == 4)193        #expect(throws: StatError.self) { try HierarchicalClustering.linkage(Matrix([[1]])) }194        #expect(throws: StatError.self) { try r.labels(clusters: 9) }195    }196}197
Files main
Menu

Theme