The optimization framework in FlinkML is a developer-oriented package that can be used to solve optimization problems common in Machine Learning (ML) tasks. In the supervised learning context, this usually involves finding a model, as defined by a set of parameters $w$, that minimize a function $f(\wv)$ given a set of $(\x, y)$ examples, where $\x$ is a feature vector and $y$ is a real number, which can represent either a real value in the regression case, or a class label in the classification case. In supervised learning, the function to be minimized is usually of the form:
\begin{equation} \label{eq:objectiveFunc} f(\wv) := \frac1n \sum_{i=1}^n L(\wv;\x_i,y_i) + \lambda\, R(\wv) \ . \end{equation}
where $L$ is the loss function and $R(\wv)$ the regularization penalty. We use $L$ to measure how well the model fits the observed data, and we use $R$ in order to impose a complexity cost to the model, with $\lambda > 0$ being the regularization parameter.
In supervised learning, we use loss functions in order to measure the model fit, by penalizing errors in the predictions $p$ made by the model compared to the true $y$ for each example. Different loss functions can be used for regression (e.g. Squared Loss) and classification (e.g. Hinge Loss) tasks.
Some common loss functions are:
Regularization in machine learning imposes penalties to the estimated models, in order to reduce overfitting. The most common penalties are the $L_1$ and $L_2$ penalties, defined as:
The $L_2$ penalty penalizes large weights, favoring solutions with more small weights rather than few large ones. The $L_1$ penalty can be used to drive a number of the solution coefficients to 0, thereby producing sparse solutions. The regularization constant $\lambda$ in $\eqref{eq:objectiveFunc}$ determines the amount of regularization applied to the model, and is usually determined through model cross-validation. A good comparison of regularization types can be found in this paper by Andrew Ng. Which regularization type is supported depends on the actually used optimization algorithm.
In order to find a (local) minimum of a function, Gradient Descent methods take steps in the direction opposite to the gradient of the function $\eqref{eq:objectiveFunc}$ taken with respect to the current parameters (weights). In order to compute the exact gradient we need to perform one pass through all the points in a dataset, making the process computationally expensive. An alternative is Stochastic Gradient Descent (SGD) where at each iteration we sample one point from the complete dataset and update the parameters for each point, in an online manner.
In mini-batch SGD we instead sample random subsets of the dataset, and compute the gradient over each batch. At each iteration of the algorithm we update the weights once, based on the average of the gradients computed from each mini-batch.
An important parameter is the learning rate $\eta$, or step size, which can be determined by one of five methods, listed below. The setting of the initial step size can significantly affect the performance of the algorithm. For some practical tips on tuning SGD see Leon Botou’s “Stochastic Gradient Descent Tricks”.
The current implementation of SGD uses the whole partition, making it effectively a batch gradient descent. Once a sampling operator has been introduced in Flink, true mini-batch SGD will be performed.
The stochastic gradient descent implementation can be controlled by the following parameters:
Parameter | Description |
---|---|
RegularizationPenalty |
The regularization function to apply. (Default value: NoRegularization) |
RegularizationConstant |
The amount of regularization to apply. (Default value: 0.1) |
LossFunction |
The loss function to be optimized. (Default value: None) |
Iterations |
The maximum number of iterations. (Default value: 10) |
LearningRate |
Initial learning rate for the gradient descent method. This value controls how far the gradient descent method moves in the opposite direction of the gradient. (Default value: 0.1) |
ConvergenceThreshold |
When set, iterations stop if the relative change in the value of the objective function $\eqref{eq:objectiveFunc}$ is less than the provided threshold, $\tau$. The convergence criterion is defined as follows: $\left| \frac{f(\wv)_{i-1} - f(\wv)_i}{f(\wv)_{i-1}}\right| < \tau$. (Default value: None) |
LearningRateMethod |
(Default value: LearningRateMethod.Default) |
Decay |
(Default value: 0.0) |
FlinkML supports Stochastic Gradient Descent with L1, L2 and no regularization. The regularization type has to implement the RegularizationPenalty
interface,
which calculates the new weights based on the gradient and regularization type.
The following list contains the supported regularization functions.
Class Name | Regularization function $R(\wv)$ |
---|---|
NoRegularization | $R(\wv) = 0$ |
L1Regularization | $R(\wv) = \norm{\wv}_1$ |
L2Regularization | $R(\wv) = \frac{1}{2}\norm{\wv}_2^2$ |
The loss function which is minimized has to implement the LossFunction
interface, which defines methods to compute the loss and the gradient of it.
Either one defines ones own LossFunction
or one uses the GenericLossFunction
class which constructs the loss function from an outer loss function and a prediction function.
An example can be seen here
val lossFunction = GenericLossFunction(SquaredLoss, LinearPrediction)
The full list of supported outer loss functions can be found here. The full list of supported prediction functions can be found here.
Function Name | Description | Loss | Loss Derivative |
---|---|---|---|
SquaredLoss |
Loss function most commonly used for regression tasks. |
$\frac{1}{2} (\wv^T \cdot \x - y)^2$ | $\wv^T \cdot \x - y$ |
LogisticLoss |
Loss function used for classification tasks. |
$\log\left(1+\exp\left( -y ~ \wv^T \cdot \x\right)\right), \quad y \in \{-1, +1\}$ | $\frac{-y}{1+\exp\left(y ~ \wv^T \cdot \x\right)}$ |
HingeLoss |
Loss function used for classification tasks. |
$\max \left(0, 1 - y ~ \wv^T \cdot \x\right), \quad y \in \{-1, +1\}$ | $\begin{cases} -y&\text{if } y ~ \wv^T <= 1 \\ 0&\text{if } y ~ \wv^T > 1 \end{cases}$ |
Function Name | Description | Prediction | Prediction Gradient |
---|---|---|---|
LinearPrediction |
The function most commonly used for linear models, such as linear regression and linear classifiers. |
$\x^T \cdot \wv$ | $\x$ |
Where:
$j$ is the iteration number
$\eta_j$ is the step size on step $j$
$\eta_0$ is the initial step size
$\lambda$ is the regularization constant
$\tau$ is the decay constant, which causes the learning rate to be a decreasing function of $j$, that is to say as iterations increase, learning rate decreases. The exact rate of decay is function specific, see Inverse Scaling and Wei Xu’s Method (which is an extension of the Inverse Scaling method).
Function Name | Description | Function | Called As |
---|---|---|---|
Default |
The function default method used for determining the step size. This is equivalent to the inverse scaling method for $\tau$ = 0.5. This special case is kept as the default to maintain backwards compatibility. |
$\eta_j = \eta_0/\sqrt{j}$ | LearningRateMethod.Default |
Constant |
The step size is constant throughout the learning task. |
$\eta_j = \eta_0$ | LearningRateMethod.Constant |
Leon Bottou's Method |
This is the |
$\eta_j = 1 / (\lambda \cdot (t_0 + j -1)) $ | LearningRateMethod.Bottou |
Inverse Scaling |
A very common method for determining the step size. |
$\eta_j = \eta_0 / j^{\tau}$ | LearningRateMethod.InvScaling |
Wei Xu's Method |
Method proposed by Wei Xu in Towards Optimal One Pass Large Scale Learning with Averaged Stochastic Gradient Descent |
$\eta_j = \eta_0 \cdot (1+ \lambda \cdot \eta_0 \cdot j)^{-\tau} $ | LearningRateMethod.Xu |
In the Flink implementation of SGD, given a set of examples in a DataSet[LabeledVector]
and
optionally some initial weights, we can use GradientDescent.optimize()
in order to optimize
the weights for the given data.
The user can provide an initial DataSet[WeightVector]
,
which contains one WeightVector
element, or use the default weights which are all set to 0.
A WeightVector
is a container class for the weights, which separates the intercept from the
weight vector. This allows us to avoid applying regularization to the intercept.