Machine Learning

From Lingoport Wiki
Jump to: navigation, search

Machine Learning Overview

Machine Learning prediction is a Globalyzer Workbench and Globalyzer Lite feature that helps users more quickly identify the real issues in their source code. We suggest applying Machine Learning as a follow-up step to scanning with Rule Sets. It helps to determine which candidate issues are indeed i18n issues.

Installation

Prerequisites: Java 8, Python 3.6.x and H2O.ai 3.x

1. Download Python version 3.6+ from website https://www.python.org/downloads/

2. Install python and add python to the PATH environment variable

3. Go to this link http://h2o-release.s3.amazonaws.com/h2o/rel-wheeler/4/index.html and make navigate to the "INSTALL IN PYTHON" tab. The instructions there are also shown below.

 Install dependencies (prepending with `sudo` if needed):
 pip install requests
 pip install tabulate
 pip install scikit-learn
 pip install colorama
 pip install future

At the command line, copy and paste these commands one line at a time:

 pip uninstall h2o
 pip install http://h2o-release.s3.amazonaws.com/h2o/rel-wheeler/4/Python/h2o-3.16.0.4-py2.py3-none-any.whl

Success if response messages have "Successfully installed h2o-3.16.0.4"

Test1: From the command line type in python -V. It is successful if it replies with a python version like "Python 3.6.2". If the python version is 2.x then there is a problem, possibly with paths that needs to be addresssed.

Test2: On the command line, go into python. In python:

> import h2o
> h2o.init()

This should complete without errors.

Work Flow

To use Machine Learning, first create a Globalyzer project with scans in the Globalyzer Workbench. Notice that the scan results includes a Predicted value.

Scan result.png

The example shows that we are in the All Scans Issues Scan View. For this view there are 4,202 issues and all are denoted as "Predicted". Since no machine learning has been done yet, all the issues have a Prediction value of '-' (Unpredicted). The Predicted count is the sum of the following Prediction values:

  • -
  • T
  • ML NULL
  • Pending
  • ML TRUE

Not included in the Predicted count is:

  • F
  • DFP
  • ML FALSE

This can be seen in the Scan Views -> Manage Scan Views... popup. The Predictions in red are tallied in the Predicted count and the black ones are not.

All scan issues.png

At the Scan Results view, right mouse click on some issues that you determine are not real issues, and choose Mark prediction: FALSE (F) from the menu. Please mark the prediction of several issues as false before applying Machine Learning.


Mark prediction as false bigger.png


After marking the prediction of several issues as false, please select Machine Learning->GO!


Machine learning go.png


And wait for the predicting process to finish.


Machine learning results.png


GO! will predict every active issue that doesn't already have a prediction value of T, F, P, or DFP. Machine Learning will assign ML True if it believes the issue is a true issue, ML False if it believe the issue is a false issue, and ML NULL if it can't decide.

Here is a listing of all possible prediction values for issues:

  • -: Issue has not been marked by a user and Machine Learning hasn't been invoked yet
  • N: Projects that ran before 6.1 are marked as N if Machine Learning is not applied. It corredsponds to another version of "-"
  • Negative: filtered issues are predicted as Negative and used to train Machine Learning.
  • T: Marked by a user as a real issue to train Machine Learning
  • F: Marked by a user as a false issue to train Machine Learning
  • P: Marked by a user as a pending issue; used to indicate that issue has been reviewed, but undecided if real issue or not
  • DFP: Issue marked as False Positive on Dashboard and can be ignored
  • ML True: Machine learning prediction that the issue is a true issue, i.e. the issue should be refactored.
  • ML False: Machine Learning prediction that the issue is a false issue and can be ignored
  • ML NULL: Machine Learning cannot make a prediction, so must be considered a true issue


If you find that issues predicted as ML False are indeed issues, please right mouse click on the issue and select Mark prediction: TRUE(T); the next time you run GO!, Machine Learning will learn your correction. If you are not satisfied with the prediction results, please continue marking more issues as F or T, and rerun Machine Learning.

Once you are satisfied with the prediction results, the issues with a prediction value of T, P, ML True, or ML NULL are the true issues that need to be addressed. The issues with a prediction value of F, ML False, or DFP can be ignored. The suggested way to view the predicted active issues is to select Scan Views->All Predicted Active.


All predicted active.png


Using the All Predicted Active Scan View hides all the issues with F, ML False, and DFP prediction values.


All predicted active results.png



After reviewing the predictions made by Machine Learning, you can either leave them as they are (ML True, ML False, ML NULL), or you can set the ML True ones explicitly to T, the ML False ones explicitly to F, and the ML NULL ones to either T or F. The difference between ML True and T is that the former is a suggestion from Machine Learning, while the later is an explicit decision. If you leave them as suggestions, then further runs of GO! may end up with different suggestions.

Tips:

  1. View all issues, including filtered issues: One way to understand some of the Machine Learning results is to show all issues, including filtered ones. When an issue is predicted as ML False, it is easier to see why when it is surrounded by filtered issues with the same type of patterns.
  2. Scan->Search in Scan Results:
    • Search on the Prediction column for issues which are ML False. From the Search panel, you can right click on the items to change the prediction with Globalyzer->Mark Prediction: TRUE (or FALSE).
    • Search on the Prediction column for issues which are ML NULL and ML True: This will help you see which issues are predicted as true issues.
  3. Sorting: In the Scan Results, a few sorts can be useful:
    • Sort on Issue: Lots of similar issues should be treated the same, either F or T. By sorting on issues, you may see a pattern to use as a category for Machine Learning
    • Sort on File: A sequence of lines may have patterns which can be used to categorize issues quickly. Use multi-selection to accelerate the T or F categorization.
    • Sort on Prediction: Instead of 'Searching', sorting on the Prediction columns can also help you see better some of the potential categories some issues fall into.

Globalyzer Lite

To use Machine Learning when running Globalyzer Lite, you need to first use the Globalyzer Workbench and invoke Machine Learning as described above on the desired scans in the project. Then, when you export the project to Lite, choose which scans you would like to use the Machine Learning.

Lite export ml.png

This sets flags in the generated project definition file (PDF) to use Machine Learning.

When Lite runs a PDF, it first scan the source using rules in the rule set. Then Lite applies the Prediction Report for the scan, if it exists. Finally, if the scan has been configured to use Machine Learning, Lite applies the Machine Learning files for the scan. The generated XML report will include Machine Learning prediction information, which will be read by the Dashboard. The Dashboard will only display Active issues with a prediction value of -, T, P, ML True, or ML NULL. Active issues with a prediction value of F, ML False, or DFP will not be displayed on the Dashboard.

Prediction Reports

Prediction Reports save prediction markings you made for a scan. When you select Machine Learning->GO!, Globalyzer automatically generates both the Prediction Report and Machine Learning files for the scan. If you prefer, you can use the Prediction Report without Machine Learning, allowing you to save and share your prediction markings without Machine Learning making predictions. Select Machine Learning->Export Prediction Report to export the Prediction Report for a scan; select Machine Learning->Import Prediction Report to update your scan results with the saved markings.


Export prediction report.png


Prediction Reports are stored in the <project>/lingoport directory. Lite will automatically use the Prediction Report, if it exists, when scanning. The Dashboard knows to ignore issues with F and ML False (if using Machine Learning) prediction values. In this way, the Workbench and Dashboard will see the same active results.

Continuous Integration System

When using Machine Learning in our Continuous Integration System, make sure everything in the project/lingoport directory is pushed to your repository. This ensures that everything downstream will work as intended.

Machine Learning FAQ

1. If I change the status of an issue, will Machine Learning work?

Yes, it will work. When you change the status of an issue, the prediction of the issue will be set by default. For example, if you move an issue to ToDo, the prediction will be marked as True; if you move an issue to Ignore/Invalid, the prediction will be marked as False. You may still manually mark the prediction of any issue, overriding the default. Then, when Machine Learning is invoked, it will use the True and False values as input to make predictions on the remaining Active issues.

Please note that as of our 6.1.1 release, we are deprecating the Invalid, Ignore, and ToDo statuses. Please mark the prediction of an issue rather than change its status.

2. In Scan Results, what is the meaning of the Predicted count?

With the introduction of Machine Learning, Scan Results has a new count display, called Predicted. This is the number of Active issues with a prediction value of -, T, P, ML True, or ML NULL. This is the number of Active issues that will appear on the Dashboard. Note that Active issues with a prediction value of F, ML False, or DFP will not be displayed on the Dashboard.

3. How does Machine Learning work?

We use h2o.ai to analyze the issue, the issue code line, and the issue reason. Based on filtered issues and your marked false issues, Machine Learning will try to find similar Active issues and set their prediction to ML False. Machine Learning prediction may be different per invocation; you won't have the exact same results every time. In addition, Machine Learning needs input to learn from, so if you only mark one issue as False, Machine Learning may not be able to find other similar issues.

4. What kind of Machine Learning algorithm does Globalyzer use?

Globalyzer uses the Gradient Boosting Machine (GBM) algorithm. Gradient Boosting Machine (for Regression and Classification) is a forward learning ensemble method. The guiding heuristic is that good predictive results can be obtained through increasingly refined approximations. H2O’s GBM sequentially builds regression trees on all the features of the dataset in a fully distributed way - each tree is built in parallel.

5. What is H2O.ai? Do I have to install it?

H2O.ai is an open source, in-memory, distributed, fast, and scalable machine learning and predictive analytics platform that allows you to build machine learning models on big data and provides easy productionalization of those models in an enterprise environment. And yes, to use Machine Learning, you must install H2O.ai to your system. It's an in-memory platform so you don't need to worry about the security of your code and data.