DecisionTreeClassifier serialization/deserialization

• I’m submitting a …

• question about how to use this project

• Summary I am trying to serialize and deserialize DecisionTreeClassifier, however it throws an error. What is a valid way to save tree to file and restore it later?

Source:

const decision = new DecisionTreeClassifier({ featureLabels: ['a','b'] });
const deserialized = new DecisionTreeClassifier({ featureLabels: ['a','b'] });

const X = [[0, 0], [1, 1]];
const y = [0, 1];
decision.fit( X, y );
const serialized = JSON.stringify(decision.toJSON())
deserialized.fromJSON(JSON.parse(serialized))

Stacktrace:

TypeError: node.question.match is not a function
    at DecisionTreeClassifier._predict (C:\apps\ml\src\lib\tree\tree.ts:434:23)
    at DecisionTreeClassifier.predict (C:\apps\ml\src\lib\tree\tree.ts:202:24)
    at Object.predict (C:\apps\ml/dttest.js:12:27)

[gilles-bara]

The problem is that JSON.stringify will convert the typed classes to regular, non-typed objects. Providing that string to a fromJSON will then only create objects with the same properties, but these objects won’t be the original classes with their appropriate methods. I suppose the fromJSON needs to be extended so that proper instances are created of the correct classes.

For the forest this could be something like this:

    BaseRandomForest.prototype.fromJSON = function (_a) {
        var _b = _a.trees, trees = _b === void 0 ? null : _b;
        if (!trees) {
            throw new Error('You must provide both tree to restore the model');
        }
        this.trees = [];
        for (const tree of trees) {
            const t = new tree_1.DecisionTreeClassifier();
            t.fromJSON(tree);
            this.trees.push(t);
        }
    };

A single classifier could then be:

    DecisionTreeClassifier.prototype.fromJSON = function (_a) {
        var _b = _a.featureLabels, featureLabels = _b === void 0 ? null : _b, _c = _a.tree, tree = _c === void 0 ? null : _c, _d = _a.verbose, verbose = _d === void 0 ? false : _d, _e = _a.random_state, random_state = _e === void 0 ? null : _e;
        this.featureLabels = featureLabels;
        const t = new DecisionNode();
        this.tree = t.fromJSON(tree);
        this.verbose = verbose;
        this.randomState = random_state;
    };

And the nodes:

    DecisionNode.prototype.fromJSON = function (_a) {
        if (typeof _a.prediction !== "undefined") {
            return Object.assign(new Leaf([]), _a);
        }
        const q = _a.question;
        this.question = new Question(q.features, q.column, q.value);
        this.falseBranch = (new DecisionNode()).fromJSON(_a.falseBranch);
        this.trueBranch = (new DecisionNode()).fromJSON(_a.trueBranch);
        return this;
    };

Allthough the fromJSON should probably not return the result to be consistent with the existing fromJSON methods, but you get the idea.

[danechung]

I have the same problem. I tried to handle using the above example, but some classes are not exported.

import { DecisionTreeClassifier } from ‘./tree’; export { DecisionTreeClassifier };

then i’m fork and fix it import { DecisionTreeClassifier, DecisionNode, Question, Leaf } from ‘./tree’; export { DecisionTreeClassifier, DecisionNode, Question, Leaf };