/**
 * 1. 为什么不做离散分割，因为使用one-hot encoding之后，按连续变量处理，可以获得不一定按照维度成员分割的效果，
 * 某一个维度可能个别成员作为单独的分割，其他则作为一个整体，这是直接写离散型分割比较难实现的。
 * 2. 多分割本质上可以被二分替代，所以没必要去做更复杂的多分割，目前也没有依据二者表现会有显著的差别。这样我们就可以直接使用二叉树来做。
 */
import { ClassifierBase } from "../base";
import { Params } from "../base/estimator";
import { SubsetSizeOption } from "../utils/paramResolvers";
export type IFeatureSplitType = 'continuous' | 'discrete';
export interface ISlice {
    X: number[][];
    Y: number[];
}
export interface IDTree {
    splitIndex: number;
    nodeValue: number;
    y: number;
    leftChild: IDTree | null;
    rightChild: IDTree | null;
    /** sample-weighted impurity decrease contributed by this split */
    weightedImpurityDecrease?: number;
    /** Class probabilities at this node, ordered by the fitted tree classes. */
    classProbabilities?: number[];
}
export interface DecisionTreeProps {
    max_depth?: number;
    min_samples_split?: number;
    criterion?: 'entropy' | 'gini';
    max_features?: SubsetSizeOption;
    randomState?: number;
}
export declare class DecisionTreeClassifier extends ClassifierBase {
    private dtree;
    private max_depth;
    private feature_number;
    private min_samples_split;
    private criterion;
    private max_features?;
    private randomState?;
    private random;
    constructor(props?: DecisionTreeProps);
    getParams(): Params;
    /** Impurity of a class-frequency vector, resolved from `criterion`. */
    private impurity;
    private selectedFeatureIndices;
    /**
     *
     * @param sampleX
     * @param sampleY
     * @param attributes indices of attributes.
     */
    treeGenerate(tree: IDTree, sampleX: number[][], sampleY: number[], depth: number): void;
    /**
     * Maximizes impurity gain over midpoint thresholds, scanning each feature
     * once in sorted order with incremental class counts.
     */
    private attributeSelection;
    private nodeImpurity;
    fit(sampleX: number[][], sampleY: number[]): void;
    predict(sampleX: number[][]): number[];
    get featureImportances(): number[];
    getRawFeatureImportances(): number[];
    private findSample;
}