fastcit
    Preparing search index...

    fastcit

    FastCIT: Fast Conditional Independence Tests (JavaScript/WASM)

    WebAssembly bindings for the data-bound conditional-independence API.

    Build the Node.js package from the repository root:

    wasm-pack build crates/ci-js --target nodejs
    
    import { Dataset, PearsonCorrelation } from "./pkg/fastcit.js";

    const data = new Dataset({
    X: { kind: "continuous", values: [0.2, 0.8, 1.4, 2.1, 2.7] },
    Y: { kind: "continuous", values: [1.0, 1.7, 2.5, 3.0, 3.8] },
    });
    const result = new PearsonCorrelation(data).runTest("X", "Y", []);
    import { Dataset, HSIC, KCI } from "./pkg/fastcit.js";

    const hsic = new HSIC(data, {
    kernel: { type: "rbf", lengthScale: 1.0 }, // or `kernels: { X: {...}, Y: {...} }`
    bandwidth: "median", // "heuristic" (default) | "median"
    nullDist: "permutation", // "gamma" (default) | "permutation"
    nPermutations: 200, // positive safe integer (default 100)
    seed: 7, // safe integer >= 0 (default 0)
    });
    hsic.runTest("X", "Y", []); // HSIC takes no conditioning set

    const kci = new KCI(data, { epsilon: 1e-3 });
    kci.runTest("X", "Y", ["Z"]);

    kernel and kernels are mutually exclusive; invalid options throw.

    import { Dataset, GCM, PillaiTrace } from "./pkg/fastcit.js";

    new GCM(data).runTest("X", "Y", ["Z"]); // continuous columns only
    new PillaiTrace(data).runTest("X", "Y", ["Z"]); // mixed discrete/continuous

    Neither class takes options (unknown keys throw). Both residualize with linear regression (pgmpy's Pillai default is a random forest). In PillaiTrace, conditioning variables enter as one-hot columns minus one level; a categorical X or Y becomes K-1 level indicators, residualized with a linear probability model. Canonical correlations are computed exactly. Degenerate inputs throw instead of returning NaN (both tests). Conditioning sets whose columns are linearly dependent (duplicate or nested categorical variables, a constant column) raise a degenerate-data error, whereas pgmpy fits them by least squares. When Z is not empty and n - columns == max(p, q), PillaiTrace returns p = 0 ("dependent"), as pgmpy does. When n - columns < max(p, q), it throws. GCM is continuous-only; a discrete X, Y or Z throws.

    Run the JavaScript checks from the package directory:

    npm ci
    npm test

    For browser builds, the complete API, and contribution guidance, see the root README and CONTRIBUTING.md.