[SPARK-20988][ML] Logistic regression uses aggregator hierarchy

mllib/src/main/scala/org/apache/spark/ml/optim/aggregator/LogisticAggregator.scala

@@ -0,0 +1,364 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements.  See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License.  You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+package org.apache.spark.ml.optim.aggregator
+
+import org.apache.spark.broadcast.Broadcast
+import org.apache.spark.internal.Logging
+import org.apache.spark.ml.feature.Instance
+import org.apache.spark.ml.linalg.{DenseVector, Vector}
+import org.apache.spark.mllib.util.MLUtils
+
+/**
+ * LogisticAggregator computes the gradient and loss for binary or multinomial logistic (softmax)
+ * loss function, as used in classification for instances in sparse or dense vector in an online
+ * fashion.
+ *
+ * Two LogisticAggregators can be merged together to have a summary of loss and gradient of
+ * the corresponding joint dataset.
+ *
+ * For improving the convergence rate during the optimization process and also to prevent against
+ * features with very large variances exerting an overly large influence during model training,
+ * packages like R's GLMNET perform the scaling to unit variance and remove the mean in order to
+ * reduce the condition number. The model is then trained in this scaled space, but returns the
+ * coefficients in the original scale. See page 9 in
+ * http://cran.r-project.org/web/packages/glmnet/glmnet.pdf
+ *
+ * However, we don't want to apply the [[org.apache.spark.ml.feature.StandardScaler]] on the
+ * training dataset, and then cache the standardized dataset since it will create a lot of overhead.
+ * As a result, we perform the scaling implicitly when we compute the objective function (though
+ * we do not subtract the mean).
+ *
+ * Note that there is a difference between multinomial (softmax) and binary loss. The binary case
+ * uses one outcome class as a "pivot" and regresses the other class against the pivot. In the
+ * multinomial case, the softmax loss function is used to model each class probability
+ * independently. Using softmax loss produces `K` sets of coefficients, while using a pivot class
+ * produces `K - 1` sets of coefficients (a single coefficient vector in the binary case). In the
+ * binary case, we can say that the coefficients are shared between the positive and negative
+ * classes. When regularization is applied, multinomial (softmax) loss will produce a result
+ * different from binary loss since the positive and negative don't share the coefficients while the
+ * binary regression shares the coefficients between positive and negative.
+ *
+ * The following is a mathematical derivation for the multinomial (softmax) loss.
+ *
+ * The probability of the multinomial outcome $y$ taking on any of the K possible outcomes is:
+ *
+ * <blockquote>
+ *    $$
+ *    P(y_i=0|\vec{x}_i, \beta) = \frac{e^{\vec{x}_i^T \vec{\beta}_0}}{\sum_{k=0}^{K-1}
+ *       e^{\vec{x}_i^T \vec{\beta}_k}} \\
+ *    P(y_i=1|\vec{x}_i, \beta) = \frac{e^{\vec{x}_i^T \vec{\beta}_1}}{\sum_{k=0}^{K-1}
+ *       e^{\vec{x}_i^T \vec{\beta}_k}}\\
+ *    P(y_i=K-1|\vec{x}_i, \beta) = \frac{e^{\vec{x}_i^T \vec{\beta}_{K-1}}\,}{\sum_{k=0}^{K-1}
+ *       e^{\vec{x}_i^T \vec{\beta}_k}}
+ *    $$
+ * </blockquote>
+ *
+ * The model coefficients $\beta = (\beta_0, \beta_1, \beta_2, ..., \beta_{K-1})$ become a matrix
+ * which has dimension of $K \times (N+1)$ if the intercepts are added. If the intercepts are not
+ * added, the dimension will be $K \times N$.
+ *
+ * Note that the coefficients in the model above lack identifiability. That is, any constant scalar
+ * can be added to all of the coefficients and the probabilities remain the same.
+ *
+ * <blockquote>
+ *    $$
+ *    \begin{align}
+ *    \frac{e^{\vec{x}_i^T \left(\vec{\beta}_0 + \vec{c}\right)}}{\sum_{k=0}^{K-1}
+ *       e^{\vec{x}_i^T \left(\vec{\beta}_k + \vec{c}\right)}}
+ *    = \frac{e^{\vec{x}_i^T \vec{\beta}_0}e^{\vec{x}_i^T \vec{c}}\,}{e^{\vec{x}_i^T \vec{c}}
+ *       \sum_{k=0}^{K-1} e^{\vec{x}_i^T \vec{\beta}_k}}
+ *    = \frac{e^{\vec{x}_i^T \vec{\beta}_0}}{\sum_{k=0}^{K-1} e^{\vec{x}_i^T \vec{\beta}_k}}
+ *    \end{align}
+ *    $$
+ * </blockquote>
+ *
+ * However, when regularization is added to the loss function, the coefficients are indeed
+ * identifiable because there is only one set of coefficients which minimizes the regularization
+ * term. When no regularization is applied, we choose the coefficients with the minimum L2
+ * penalty for consistency and reproducibility. For further discussion see:
+ *
+ * Friedman, et al. "Regularization Paths for Generalized Linear Models via Coordinate Descent"
+ *
+ * The loss of objective function for a single instance of data (we do not include the
+ * regularization term here for simplicity) can be written as
+ *
+ * <blockquote>
+ *    $$
+ *    \begin{align}
+ *    \ell\left(\beta, x_i\right) &= -log{P\left(y_i \middle| \vec{x}_i, \beta\right)} \\
+ *    &= log\left(\sum_{k=0}^{K-1}e^{\vec{x}_i^T \vec{\beta}_k}\right) - \vec{x}_i^T \vec{\beta}_y\\
+ *    &= log\left(\sum_{k=0}^{K-1} e^{margins_k}\right) - margins_y
+ *    \end{align}
+ *    $$
+ * </blockquote>
+ *
+ * where ${margins}_k = \vec{x}_i^T \vec{\beta}_k$.
+ *
+ * For optimization, we have to calculate the first derivative of the loss function, and a simple
+ * calculation shows that
+ *
+ * <blockquote>
+ *    $$
+ *    \begin{align}
+ *    \frac{\partial \ell(\beta, \vec{x}_i, w_i)}{\partial \beta_{j, k}}
+ *    &= x_{i,j} \cdot w_i \cdot \left(\frac{e^{\vec{x}_i \cdot \vec{\beta}_k}}{\sum_{k'=0}^{K-1}
+ *      e^{\vec{x}_i \cdot \vec{\beta}_{k'}}\,} - I_{y=k}\right) \\
+ *    &= x_{i, j} \cdot w_i \cdot multiplier_k
+ *    \end{align}
+ *    $$
+ * </blockquote>
+ *
+ * where $w_i$ is the sample weight, $I_{y=k}$ is an indicator function
+ *
+ *  <blockquote>
+ *    $$
+ *    I_{y=k} = \begin{cases}
+ *          1 & y = k \\
+ *          0 & else
+ *       \end{cases}
+ *    $$
+ * </blockquote>
+ *
+ * and
+ *
+ * <blockquote>
+ *    $$
+ *    multiplier_k = \left(\frac{e^{\vec{x}_i \cdot \vec{\beta}_k}}{\sum_{k=0}^{K-1}
+ *       e^{\vec{x}_i \cdot \vec{\beta}_k}} - I_{y=k}\right)
+ *    $$
+ * </blockquote>
+ *
+ * If any of margins is larger than 709.78, the numerical computation of multiplier and loss
+ * function will suffer from arithmetic overflow. This issue occurs when there are outliers in
+ * data which are far away from the hyperplane, and this will cause the failing of training once
+ * infinity is introduced. Note that this is only a concern when max(margins) &gt; 0.
+ *
+ * Fortunately, when max(margins) = maxMargin &gt; 0, the loss function and the multiplier can
+ * easily be rewritten into the following equivalent numerically stable formula.
+ *
+ * <blockquote>
+ *    $$
+ *    \ell\left(\beta, x\right) = log\left(\sum_{k=0}^{K-1} e^{margins_k - maxMargin}\right) -
+ *       margins_{y} + maxMargin
+ *    $$
+ * </blockquote>
+ *
+ * Note that each term, $(margins_k - maxMargin)$ in the exponential is no greater than zero; as a
+ * result, overflow will not happen with this formula.
+ *
+ * For $multiplier$, a similar trick can be applied as the following,
+ *
+ * <blockquote>
+ *    $$
+ *    multiplier_k = \left(\frac{e^{\vec{x}_i \cdot \vec{\beta}_k - maxMargin}}{\sum_{k'=0}^{K-1}
+ *       e^{\vec{x}_i \cdot \vec{\beta}_{k'} - maxMargin}} - I_{y=k}\right)
+ *    $$
+ * </blockquote>
+ *
+ *
+ * @param bcCoefficients The broadcast coefficients corresponding to the features.
+ * @param bcFeaturesStd The broadcast standard deviation values of the features.
+ * @param numClasses the number of possible outcomes for k classes classification problem in
+ *                   Multinomial Logistic Regression.
+ * @param fitIntercept Whether to fit an intercept term.
+ * @param multinomial Whether to use multinomial (softmax) or binary loss
+ * @note In order to avoid unnecessary computation during calculation of the gradient updates
+ * we lay out the coefficients in column major order during training. This allows us to
+ * perform feature standardization once, while still retaining sequential memory access
+ * for speed. We convert back to row major order when we create the model,
+ * since this form is optimal for the matrix operations used for prediction.
+ */
+private[ml] class LogisticAggregator(
+    bcFeaturesStd: Broadcast[Array[Double]],
+    numClasses: Int,
+    fitIntercept: Boolean,
+    multinomial: Boolean)(bcCoefficients: Broadcast[Vector])
+  extends DifferentiableLossAggregator[Instance, LogisticAggregator] with Logging {
+
+  private val numFeatures = bcFeaturesStd.value.length
+  private val numFeaturesPlusIntercept = if (fitIntercept) numFeatures + 1 else numFeatures
+  private val coefficientSize = bcCoefficients.value.size
+  protected override val dim: Int = coefficientSize
+  if (multinomial) {
+    require(numClasses ==  coefficientSize / numFeaturesPlusIntercept, s"The number of " +
+      s"coefficients should be ${numClasses * numFeaturesPlusIntercept} but was $coefficientSize")
+  } else {
+    require(coefficientSize == numFeaturesPlusIntercept, s"Expected $numFeaturesPlusIntercept " +
+      s"coefficients but got $coefficientSize")
+    require(numClasses == 1 || numClasses == 2, s"Binary logistic aggregator requires numClasses " +
+      s"in {1, 2} but found $numClasses.")
+  }
+
+  @transient private lazy val coefficientsArray: Array[Double] = bcCoefficients.value match {
+    case DenseVector(values) => values
+    case _ => throw new IllegalArgumentException(s"coefficients only supports dense vector but " +
+      s"got type ${bcCoefficients.value.getClass}.)")
+  }
+
+  if (multinomial && numClasses <= 2) {
+    logInfo(s"Multinomial logistic regression for binary classification yields separate " +
+      s"coefficients for positive and negative classes. When no regularization is applied, the" +
+      s"result will be effectively the same as binary logistic regression. When regularization" +
+      s"is applied, multinomial loss will produce a result different from binary loss.")
+  }
+
+  /** Update gradient and loss using binary loss function. */
+  private def binaryUpdateInPlace(features: Vector, weight: Double, label: Double): Unit = {
+
+    val localFeaturesStd = bcFeaturesStd.value
+    val localCoefficients = coefficientsArray
+    val localGradientArray = gradientSumArray
+    val margin = - {
+      var sum = 0.0
+      features.foreachActive { (index, value) =>
+        if (localFeaturesStd(index) != 0.0 && value != 0.0) {
+          sum += localCoefficients(index) * value / localFeaturesStd(index)
+        }
+      }
+      if (fitIntercept) sum += localCoefficients(numFeaturesPlusIntercept - 1)
+      sum
+    }
+
+    val multiplier = weight * (1.0 / (1.0 + math.exp(margin)) - label)
+
+    features.foreachActive { (index, value) =>
+      if (localFeaturesStd(index) != 0.0 && value != 0.0) {
+        localGradientArray(index) += multiplier * value / localFeaturesStd(index)
+      }
+    }
+
+    if (fitIntercept) {
+      localGradientArray(numFeaturesPlusIntercept - 1) += multiplier
+    }
+
+    if (label > 0) {
+      // The following is equivalent to log(1 + exp(margin)) but more numerically stable.
+      lossSum += weight * MLUtils.log1pExp(margin)
+    } else {
+      lossSum += weight * (MLUtils.log1pExp(margin) - margin)
+    }
+  }
+
+  /** Update gradient and loss using multinomial (softmax) loss function. */
+  private def multinomialUpdateInPlace(features: Vector, weight: Double, label: Double): Unit = {
+    // TODO: use level 2 BLAS operations
+    /*
+      Note: this can still be used when numClasses = 2 for binary
+      logistic regression without pivoting.
+     */
+    val localFeaturesStd = bcFeaturesStd.value
+    val localCoefficients = coefficientsArray
+    val localGradientArray = gradientSumArray
+
+    // marginOfLabel is margins(label) in the formula
+    var marginOfLabel = 0.0
+    var maxMargin = Double.NegativeInfinity
+
+    val margins = new Array[Double](numClasses)
+    features.foreachActive { (index, value) =>
+      val stdValue = value / localFeaturesStd(index)
+      var j = 0
+      while (j < numClasses) {
+        margins(j) += localCoefficients(index * numClasses + j) * stdValue
+        j += 1
+      }
+    }
+    var i = 0
+    while (i < numClasses) {
+      if (fitIntercept) {
+        margins(i) += localCoefficients(numClasses * numFeatures + i)
+      }
+      if (i == label.toInt) marginOfLabel = margins(i)
+      if (margins(i) > maxMargin) {
+        maxMargin = margins(i)
+      }
+      i += 1
+    }
+
+    /**
+     * When maxMargin is greater than 0, the original formula could cause overflow.
+     * We address this by subtracting maxMargin from all the margins, so it's guaranteed
+     * that all of the new margins will be smaller than zero to prevent arithmetic overflow.
+     */
+    val multipliers = new Array[Double](numClasses)
+    val sum = {
+      var temp = 0.0
+      var i = 0
+      while (i < numClasses) {
+        if (maxMargin > 0) margins(i) -= maxMargin
+        val exp = math.exp(margins(i))
+        temp += exp
+        multipliers(i) = exp
+        i += 1
+      }
+      temp
+    }
+
+    margins.indices.foreach { i =>
+      multipliers(i) = multipliers(i) / sum - (if (label == i) 1.0 else 0.0)
+    }
+    features.foreachActive { (index, value) =>
+      if (localFeaturesStd(index) != 0.0 && value != 0.0) {
+        val stdValue = value / localFeaturesStd(index)
+        var j = 0
+        while (j < numClasses) {
+          localGradientArray(index * numClasses + j) += weight * multipliers(j) * stdValue
+          j += 1
+        }
+      }
+    }
+    if (fitIntercept) {
+      var i = 0
+      while (i < numClasses) {
+        localGradientArray(numFeatures * numClasses + i) += weight * multipliers(i)
+        i += 1
+      }
+    }
+
+    val loss = if (maxMargin > 0) {
+      math.log(sum) - marginOfLabel + maxMargin
+    } else {
+      math.log(sum) - marginOfLabel
+    }
+    lossSum += weight * loss
+  }
+
+  /**
+   * Add a new training instance to this LogisticAggregator, and update the loss and gradient
+   * of the objective function.
+   *
+   * @param instance The instance of data point to be added.
+   * @return This LogisticAggregator object.
+   */
+  def add(instance: Instance): this.type = {
+    instance match { case Instance(label, weight, features) =>
+      require(numFeatures == features.size, s"Dimensions mismatch when adding new instance." +
+        s" Expecting $numFeatures but got ${features.size}.")
+      require(weight >= 0.0, s"instance weight, $weight has to be >= 0.0")
+
+      if (weight == 0.0) return this
+
+      if (multinomial) {
+        multinomialUpdateInPlace(features, weight, label)
+      } else {
+        binaryUpdateInPlace(features, weight, label)
+      }
+      weightSum += weight
+      this
+    }
+  }
+}