Compare DataFrames

core/src/main/kotlin/org/jetbrains/kotlinx/dataframe/impl/api/compareDataFrames.kt

@@ -0,0 +1,224 @@
+package org.jetbrains.kotlinx.dataframe.impl.api
+
+import org.jetbrains.kotlinx.dataframe.DataFrame
+import org.jetbrains.kotlinx.dataframe.DataRow
+import org.jetbrains.kotlinx.dataframe.annotations.DataSchema
+import org.jetbrains.kotlinx.dataframe.api.DataRowSchema
+import org.jetbrains.kotlinx.dataframe.api.concat
+import org.jetbrains.kotlinx.dataframe.api.dataFrameOf
+import org.jetbrains.kotlinx.dataframe.api.emptyDataFrame
+import org.jetbrains.kotlinx.dataframe.nrow
+
+@DataSchema
+internal class ComparisonDescription<T>(
+    val rowAtIndex: Int,
+    val of: DataFrameOfComparison,
+    val modification: RowOfComparison,
+    val insertedAfterRow: Int?,
+    val modifiedRowContent: DataRow<T>,
+) : DataRowSchema
+
+internal enum class DataFrameOfComparison {
+    DFA,
+    DFB,
+}
+
+internal enum class RowOfComparison {
+    INSERTED,
+    REMOVED,
+}
+
+/**
+ * Returns a DataFrame whose rows explain the differences between dfA and dfB.
+ * One must think of the set of commands in a script as being executed simultaneously
+ */
+internal fun <T> compareDataFramesImpl(dfA: DataFrame<T>, dfB: DataFrame<T>): DataFrame<ComparisonDescription<T>> {
+    var comparisonDf = emptyDataFrame<ComparisonDescription<T>>()
+    // compare by exploiting Myers difference algorithm
+    val shortestEditScript = myersDifferenceAlgorithmImpl(dfA, dfB)
+    for (i in 1 until shortestEditScript.size) {
+        val x = shortestEditScript[i].first
+        val y = shortestEditScript[i].second
+        val xPrev = shortestEditScript[i - 1].first
+        val yPrev = shortestEditScript[i - 1].second
+        when {
+            // row at index 'x-1' of dfA was removed
+            xPrev + 1 == x && yPrev + 1 != y -> {
+                val indexOfRemovedRow = x - 1
+                val sourceDfOfRemovedRow = DataFrameOfComparison.DFA
+                val rowContent = dfA[indexOfRemovedRow]
+                comparisonDf = comparisonDf.concat(
+                    dataFrameOf(
+                        ComparisonDescription(
+                            indexOfRemovedRow,
+                            sourceDfOfRemovedRow,
+                            RowOfComparison.REMOVED,
+                            null,
+                            rowContent,
+                        ),
+                    ),
+                )
+            }
+
+            // row at index 'y-1' of dfB was inserted after row in position 'x-1' of dfA
+            yPrev + 1 == y && xPrev + 1 != x -> {
+                val indexOfInsertedRow = y - 1
+                val sourceDfOfInsertedRow = DataFrameOfComparison.DFB
+                val indexOfReferenceRow = x - 1
+                val rowContent = dfB[indexOfInsertedRow]
+                comparisonDf = comparisonDf.concat(
+                    dataFrameOf(
+                        ComparisonDescription(
+                            indexOfInsertedRow,
+                            sourceDfOfInsertedRow,
+                            RowOfComparison.INSERTED,
+                            indexOfReferenceRow,
+                            rowContent,
+                        ),
+                    ),
+                )
+            }
+        }
+    }
+    return comparisonDf
+}
+
+/**
+ * dfs with same schema. Returns an optimal path from origin to (N,M) in the edit graph.
+ * N is dfA.nrow, M is dfB.nrow.
+ * Knowing this path is knowing the differences between dfA and dfB
+ * and the shortest edit script to get B from A.
+ * The cost of this alg's worst case in O( (N+M)D ), D is the length of shortest edit script.
+ *
+ * The idea of the algorithm is the following: try to cross the edit graph making 'd' non-diagonal moves,
+ * increase 'd' until you succeed.
+ * Non-diagonal moves make edit script longer, while diagonal moves do not.
+ *
+ * snake: non-diagonal edge and then a possibly empty sequence of diagonal edges
+ * D-path: a path starting at (0,0) that has exactly D non-diagonal edges
+ */
+internal fun <T> myersDifferenceAlgorithmImpl(dfA: DataFrame<T>, dfB: DataFrame<T>): List<Pair<Int, Int>> {
+    // Return value
+    val path = mutableListOf<Pair<Int, Int>>()
+    // 'ses' stands for shortest edit script, next var is never returned, it is in the code
+    // to show the capabilities of the algorithm
+    var sesLength: Int
+    val sumOfLength = dfA.nrow + dfB.nrow
+    // matrix containing the endpoint of the furthest reaching D-path ending in diagonal k
+    // for each d-k couple of interest
+    val v = mutableListOf<IntArray>()
+    repeat(sumOfLength + 1) {
+        v.add(IntArray(sumOfLength * 2 + 1))
+    }
+    var isOver = false
+    // starting the algorithm
+    // 0 position is -(M+N) position in the alg's paper -> need to normalize each access to v
+    val normalizer = sumOfLength
+    v[0][1 + normalizer] = 0 // fictitious
+    // d is the number of non-diagonal edges
+    var d = 0
+    while (d <= sumOfLength && !isOver) {
+        for (k in -d..d step 2) {
+            var x: Int
+            // Each furthest reaching D-path ending in diagonal k
+            // is built by exploiting the furthest reaching (D-1)-path ending in k-1 or (exclusive or) k+1
+            if (k == -d || k != d && v[d][k - 1 + normalizer] < v[d][k + 1 + normalizer]) {
+                x = v[d][k + 1 + normalizer]
+            } else {
+                x = v[d][k - 1 + normalizer] + 1
+            }
+            var y = x - k
+            while (x < dfA.nrow && y < dfB.nrow && dfA[x] == dfB[y]) {
+                x += 1
+                y += 1
+            }
+            v[d][k + normalizer] = x
+            // need this data in the next iteration
+            if (d < sumOfLength) {
+                v[d + 1][k + normalizer] = x
+            }
+            // Edit graph was fully crossed
+            if (x >= dfA.nrow && y >= dfB.nrow) {
+                isOver = true
+                sesLength = d
+                recursivePathFill(path, v, d, k, normalizer, dfA, dfB)
+                break
+            }
+        }
+        // try with a longer edit script
+        d++
+    }
+    val immutablePath = path.toList()
+    return immutablePath
+}
+
+internal tailrec fun <T> recursivePathFill(
+    path: MutableList<Pair<Int, Int>>,
+    v: MutableList<IntArray>,
+    d: Int,
+    k: Int,
+    normalizer: Int,
+    dfA: DataFrame<T>,
+    dfB: DataFrame<T>,
+) {
+    // Enlist my self
+    val xCurrent = v[d][k + normalizer]
+    val yCurrent = xCurrent - k
+    path.add(Pair(xCurrent, yCurrent))
+    // I look for endpoint I was built from, it is represented by kPrev.
+    // It will be an argument of the next recursive step.
+    // Moreover, I need to enlist the points composing the snake that precedes me (it may be empty).
+    if (d > 0) {
+        var kPrev: Int
+        var xSnake: Int
+        if (k == -d || k != d && v[d][k - 1 + normalizer] < v[d][k + 1 + normalizer]) {
+            kPrev = k + 1
+            xSnake = v[d - 1][kPrev + normalizer]
+        } else {
+            kPrev = k - 1
+            xSnake = v[d - 1][kPrev + normalizer] + 1
+        }
+        var ySnake = xSnake - k
+        val snake = mutableListOf<Pair<Int, Int>>()
+        do {
+            snake.add(0, Pair(xSnake, ySnake))
+            if (xSnake == xCurrent && ySnake == yCurrent) {
+                if (snake.isNotEmpty()) {
+                    snake.removeFirst()
+                    for (e in snake) {
+                        path.add(e)
+                    }
+                }
+                recursivePathFill(path, v, d - 1, kPrev, normalizer, dfA, dfB)
+                return
+            }
+            if (xSnake < dfA.nrow &&
+                ySnake < dfB.nrow &&
+                xSnake >= 0 &&
+                ySnake >= 0 &&
+                dfA[xSnake] == dfB[ySnake]
+            ) {
+                xSnake += 1
+                ySnake += 1
+            }
+        }
+        while (xSnake <= xCurrent && ySnake <= yCurrent)
+    }
+    // Step base.
+    // Eventually need to add diagonal edges from origin to the furthest reaching point with d=0.
+    // Moreover, the path is reversed so that it can be read from left to right correctly
+    if (d == 0) {
+        if (path.last().first != 0 && path.last().second != 0) {
+            val last = path.last()
+            var x = last.first - 1
+            var y = last.second - 1
+            while (x >= 0 && y >= 0) {
+                path.add(Pair(x, y))
+                x -= 1
+                y -= 1
+            }
+        }
+        path.reverse()
+        return
+    }
+}