##### Site index · List index
Message view
Top
From tillrohrmann <...@git.apache.org>
Date Tue, 17 May 2016 10:19:18 GMT
Github user tillrohrmann commented on a diff in the pull request:

--- Diff: docs/apis/batch/libs/ml/cross_validation.md ---
@@ -0,0 +1,175 @@
+---
+mathjax: include
+title: Cross Validation
+
+sub-nav-group: batch
+sub-nav-title: Cross Validation
+---
+<!--
+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
+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
+
+
+Unless required by applicable law or agreed to in writing,
+"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
+-->
+
+* This will be replaced by the TOC
+{:toc}
+
+## Description
+
+ A prevalent problem when utilizing machine learning algorithms is *overfitting*, or
when an algorithm "memorizes" the training data but does a poor job extrapolating to out of
sample cases. A common method for dealing with the overfitting problem is to hold back some
subset of data from the original training algorithm and then measure the fit algorithm's performance
on this hold-out set. This is commonly known as *cross validation*.  A model is trained on
one subset of data and then *validated* on another set of data.
+
+## Cross Validation Strategies
+
+There are several strategies for holding out data. FlinkML has convenience methods for
+- Train-Test Splits
+- Train-Test-Holdout Splits
+- K-Fold Splits
+- Multi-Random Splits
+
+### Train-Test Splits
+
+The simplest method of splitting is the trainTestSplit. This split takes a DataSet
and a parameter *fraction*.  The *fraction* indicates the portion of the DataSet that should
be allocated to the training set. This split also takes two additional optional parameters,
*precise* and *seed*.
+
+By default, the Split is done by randomly deciding weather or not an observation is assigned
to the training DataSet with probability = *fraction*.  When *precise* is true however,
additional steps are taken to ensure the training set is as close as possible to the length
of the DataSet  $\cdot$ *fraction*.
+
+The method returns a new TrainTestDataSet object which has a .training attribute
containing the training DataSet and a .testing attribute containing the testing DataSet.
+
+
+### Train-Test-Holdout Splits
+
+In some cases, algorithms have been known to 'learn' the testing set.  To combat this
issue, a train-test-hold out strategy introduces a secondary holdout set, aptly called the
*holdout* set.
+
+Traditionally, training and testing would be done to train an algorithms as normal and
then a final test of the algorithm on the holdout set would be done.  Ideally, prediction
errors/model scores in the holdout set would not be significantly different than those observed
in the testing set.
+
+In a train-test-holdout strategy we sacrifice the sample size of the initial fitting
algorithm for increased confidence that our model is not over-fit.
+
+When using trainTestHoldout splitter, the *fraction* Double is replaced by a *fraction*
array of length three. The first element coresponds to the portion to be used for training,
second for testing, and third for holdout.  The weights of this array are *relative*, e.g.
an array Array(3.0, 2.0, 1.0) would results in approximately 50% of the observations being
in the training set, 33% of the observations in the testing set, and 17% of the observations
in holdout set.
+
+### K-Fold Splits
+
+In a *k-fold* strategy, the DataSet is split into *k* equal subsets. Then for each of
the *k* subsets, a TrainTestDataSet is created where the subset is the .training DataSet,
and the remaining subsets are the .testing set.
+
+For each training set, an algorithm is trained and then is evaluated based on the predictions
based on the assosciated testing set. When an algorithm that has consistent grades (e.g. prediction
errors) across held out datasets we can have some confidence that our approach (e.g. choice
of algorithm / algorithm parameters / number of iterations) is robust against overfitting.
--- End diff --

typo: assosciated --> associated

---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastructure@apache.org or file a JIRA ticket
with INFRA.
---


Mime
View raw message