improve splits

split.ts

@@ -1,16 +1,67 @@
 import pl from "npm:nodejs-polars";
 
-export function trainTestSplit(df: pl.DataFrame, size, ...yFeatures: string[]) {
+type DfSplit = {
-  let shuffle = df.sample(df.height - 1);
+  trainX: pl.DataFrame;
-  let testSize = Math.round(shuffle.shape.height * size);
+  trainY: pl.DataFrame;
-  let trainSize = shuffle.shape.height - testSize;
+  testX: pl.DataFrame;
-  let [train, test] = [shuffle.head(trainSize), shuffle.tail(testSize)];
+  testY: pl.DataFrame;
-  let [trainY, testY] = [train.select(...yFeatures), test.select(...yFeatures)];
+  size: number;
-  let [trainX, testX] = [train.drop(...yFeatures), test.drop(...yFeatures)];
+};
-  return {
+
-    trainX,
+export function sliceK(
-    trainY,
+  df: pl.DataFrame,
-    testX,
+  size: number,
-    testY,
+  k: number,
-  };
+  ...yFeatures: string[]
+): DfSplit[] {
+  let testSize = Math.round(df.shape.height * size);
+  while (testSize % k !== 0) {
+    testSize -= 1;
+  }
+  if (df.shape.height / testSize < k) {
+    throw new Error(
+      `k value is too large, max k value is ${df.shape.height / testSize}`,
+    );
+  }
+  let trainSize = df.shape.height - testSize;
+  let result: DfSplit[] = [];
+  let data = df;
+  for (let i = 0; i < k; i++) {
+    let [train, test] = [data.head(trainSize), data.tail(testSize)];
+    let [trainY, testY] = [
+      train.select(...yFeatures),
+      test.select(...yFeatures),
+    ];
+    let [trainX, testX] = [train.drop(...yFeatures), test.drop(...yFeatures)];
+    result.push({
+      trainX,
+      trainY,
+      testX,
+      testY,
+      size,
+    });
+    data = pl.concat([test, train]);
+  }
+  return result;
+}
+
+export function trainTestSplit(
+  df: pl.DataFrame,
+  size: number,
+  shuffle = true,
+  ...yFeatures: string[]
+) {
+  let data = shuffle ? df.sample(df.height - 1) : df;
+  const result = sliceK(data, size, 1, ...yFeatures);
+  return result[0];
+}
+
+export function kFold(
+  df: pl.DataFrame,
+  k: number,
+  shuffle = true,
+  ...yFeatures: string[]
+): DfSplit[] {
+  let data = shuffle ? df.sample(df.height - 1) : df;
+  return sliceK(data, 1 / k, k, ...yFeatures);
 }