Distributed stage-wise parallel machine learning
A method for machine learning a data set in a data processing framework is disclosed. A forest is trained with the data set that generates a plurality of trees in parallel. Each tree includes leaf nodes having a constant weight. A discriminative value for each leaf node is learned with a supervised model. The forest is reconstructed with the discriminative values replacing the constant weight for each leaf node.
1. A method for machine learning a data set in a data processing framework, comprising:
generating a forest comprising a plurality of trees that each include a plurality of leaf nodes, said generating based on initial training data that includes the data set, said generating including:
generating the plurality of trees in parallel, and
assigning a constant weight to each leaf node of the plurality of leaf nodes;
constructing new training data based on the generated forest and the initial training data;
determining a discriminative value for each of the plurality of leaf nodes in parallel based on the new training data to determine a plurality of discriminative values; and
reconstructing the forest with the discriminative values replacing the constant weight for each of the plurality of leaf nodes.
2. The method of claim 1 wherein the initial training data includes over a hundred million of examples and over a thousand features.
3. The method of claim 1 wherein said generating a forest includes training with a random forest algorithm.
4. The method of claim 1 wherein said constructing comprises using a supervised model for learning the discriminative value that is a linear model.
5. The method of claim 1 wherein the discriminative value of the leaf node represents the significance of leaf node with respect to the other leaf nodes.
6. The method of claim 1 wherein the generating a forest, determining, and reconstructing are performed in sequence.
7. The method of claim 6 wherein each tree of the plurality of trees is generated separately.
8. The method of claim 6 wherein the discriminative values are determined from linear learning performed in parallel with a parallel optimization model.
9. The method of claim 8 wherein the parallel optimization model is a Limited memory Broyden-Fletcher-Goldfarb-Shanno algorithm.
10. The method of claim 6 wherein determining the discriminative values is performed with a convex algorithm.
11. A computer readable storage medium storing computer executable instructions for controlling a computing device to perform a method for machine learning a data set in a data processing framework, the method comprising:
generating a forest comprising a plurality of trees that each include a plurality of leaf nodes, said generating based on initial training data including the data set, said generating including:
generating the plurality of trees in parallel, and
assigning a constant weight to each leaf node;
constructing new training data based on the generated forest and the initial training data;
determining a discriminative value for each of the plurality of leaf nodes in parallel based on the new training data to determine a plurality of discriminative values; and
reconstructing the forest with the discriminative values replacing the constant weight for each of the plurality of leaf nodes.
12. The computer readable storage medium of claim 11 wherein the method is performed in a data processing framework configured to run on a distributed data storage and processing system.
13. The computer readable storage medium of claim 12 wherein the data processing framework is selected from a group of frameworks including a map-reduce type environment and a Structured Computations Optimized for Parallel Execution environment.
14. The computer readable storage medium of claim 11 comprising constructing new training data from the data set, and learning the discriminative value from the new training data.
15. The computer readable storage medium of claim 11 wherein the generating a forest, determining, and reconstructing are performed in sequence on the data processing framework.
16. The computer readable storage medium of claim 15 wherein each tree of the plurality of trees is generated separately on the data processing framework.
17. The computer readable storage medium of claim 15 wherein the discriminative values are determined from linear learning performed in parallel on the data processing framework.
18. A method for machine learning a data set in a data processing framework:
generating a forest comprising a plurality of trees that each include a plurality of leaf nodes, said generating based on initial training data including the data set, said generating including:
generating the plurality of trees in parallel, and
assigning a constant weight to each leaf node;
constructing new training data including a linear model;
determining a discriminative value from the new training data for each of the plurality of leaf nodes in parallel based on the new training data to determine a plurality of discriminative values; and
reconstructing the forest with the discriminative values replacing the constant weight for each of the plurality of leaf nodes.
19. The method of claim 18 , wherein constructing new training data includes determining a prediction value for each leaf node and performing a prediction on each example of the data set.
20. The method of claim 19 wherein a feature vector is constructed from each prediction value and corresponding leaf node, and the feature vector is combined with labels of the data set.