Press "Enter" to skip to content

Category: Machine Learning

LIME

Published 2016-09-02 by Kevin Feasel

William Vorhies discusses a new technical paper on Local Interpretable Model-Agnostic Explanations:

What the model actually used for classification were these: ‘posting’, ‘host’, ‘NNTP’, ‘EDU’, ‘have’, ‘there’.  These are meaningless artifacts that appear in both the training and test sets and have nothing to do with the topic except that, for example, the word “posting” (part of the email header) appears in 21.6% of the examples in the training set but only two times in the class “Christianity.”

Is this model going to generalize?  Absolutely not.

An Example from Image Processing

In this example using Google’s Inception NN on arbitrary images the objective was to correctly classify “tree frogs”.  The classifier was correct in about 54% of cases but also interpreted the image as a pool table (7%) and a balloon (5%).

Looks like an interesting paper.  Click through for a link to the paper.

Comments closed

The Joy Of Hyperparameters

Published 2016-09-01 by Kevin Feasel

Koos van Strien shows how to tune hyperparameters using Azure ML:

Today, we’ll focus on tuning the model’s properties. We won’t discuss the details of all properties (you can easily look that up in the docs), instead we’ll look at how to test for different parameter combinations insize Azure ML Studio.

As soon as you click on an untrained model inside your experiment, you’ll be presented with some parameters – or, in ML parlance, hyperparameters – you can tweak.

Parameter tuning is pretty easy using Azure ML.

Comments closed

Training Data With Azure ML

Published 2016-08-26 by Kevin Feasel

Koos van Strien discusses training data sets and cross-validating results:

When choosing a train and testset, you’ll implicitly introduce a new bias: it could be that the model you just trained predicts well for this particular testset, when trained for this particular trainset. To reduce this bias, you could “cross-validate” your results.

Cross-validation (often abbreviated as just “cv”) splits the dataset into n folds. Each fold is used once as a testset, using all other folds together as a training set. So in our pizza example with 100 records, with 5 folds we will have 5 test runs:

This isn’t Azure ML-specific, and is good reading.

Comments closed

Getting Started With Azure ML

Published 2016-08-24 by Kevin Feasel

Koos van Strien gives a quick overview of Azure ML:

Before I started, I was already quite comfortable programming Python and did some R programming in the past. This turned out pretty handy, though not really needed to start off with – because starting with Azure ML, the data flow can be created much like BI specialists are used to in SSIS.

A good place to start for me was the Tutorial competition (Iris Petal Competition). It provides you with a pre-filled workspace with everything in place to train and test your first ML model:

I’d like to see Azure ML get more traction; I’m not optimistic that it will.

Comments closed

Calling Azure ML Web Services Using Data Factory

Published 2016-08-22 by Kevin Feasel

Ginger Grant shows how to call an Azure Machine Learning web service from within Azure Data Factory:

The Linked Service for ML is going to need some information from the Web Service, the URL and the API key. Chances are neither of these have been committed to memory, instead open up Azure ML, go to Web Service and copy them. For the URL, look under the API Help Pagegrid, there are two options, Request/Response and Batch Execution. Clicking on Batch Execution loads a new page Batch Execution API Document. The URL can be found under Request URI. When copying the URL, you do not need to include any text after the word “jobs”. The rest of the URL, “?api-version=2.0”. Copying the entire URL will cause an error. Going back to the web Services page, The API Key appears on the dashboard section of Azure ML and there is a convenient button for copying it. Using these two pieces of information, it is now possible to create the Data Factory Linked Service to make the connection to the web service, which here I called AzureMLLinkedService

Read the whole thing.

Comments closed

Amazon Machine Learning

Published 2016-08-08 by Kevin Feasel

Ujjwal Ratan uses patient readmission data to demonstrate Amazon Machine Learning:

The Amazon ML endpoint created earlier can be invoked using an API call. This is very handy for building an application for end users who can interact with the ML model in real time.

Create a similar application and host it as a static website on Amazon S3. This feature of S3 allows you to host websites without any web servers and takes away the complexities of scaling hardware based on traffic routed to your application. The following is a screenshot from the application:

I think that Azure ML is still ahead of Amazon’s ML solution, but I’m happy to see the competition.

Comments closed

Azure ML Updates

Published 2016-08-03 by Kevin Feasel

David Smith walks us through new language engines supported in Azure ML:

ML studio now gives you even more flexibility, with new language engines supported in the language modules. Within the Execute Python Script module, you can now choose to use Python 2.7.11 or Python 3.5, both of which run within the Acaconda 4.0 distribution. And within the Execute R Script module, you can now choose Microsoft R Open 3.2.2 as your R engine, in addition to the existing CRAN R 3.1.0 engine. Microsoft R Open 3.2.2 not only gives you a newer R language engine, it also gives you access to a wealth of new R packages for use within ML Studio. Over 400 packages are pre-installed for use with the R Script module, and you can install and use any other R package (including CRAN packages and your own R packages) via the Script Bundle input port.

I’m interested in the Microsoft R Open language support, as Azure ML’s still using a relatively older version of R (3.1.0).

Comments closed

Approximation Or Classification?

Published 2016-07-08 by Kevin Feasel

A blog post on the Algolytics blog discusses different approximation and classification models and when to use each:

Even if your target variable is a numeric one, sometimes it’s better to use classification methods instead of approximation ones. For instance if you have mostly zero target values and just a few non-zero values. Change the latter to 1, in this case you’ll have two categories: 1 (positive value of your target variable ) and 0. You can also split numerical variable into multiple subgroups : apartment prices for low, medium and high by equal subset width and predict them using classification algorithms. This process is called discretization.

Both types of models are common in machine learning, so a good understanding of when to use which is important.

Comments closed

Running Compiled Code In Azure ML

Published 2016-07-04 by Kevin Feasel

Max Kaznady shows how to use R or Python scripts to call compiled code within Azure ML:

In this post, we focus on sourcing R and Python’s external dependencies, such as R libraries and Python modules, which are not already installed on Azure ML and require code compilation. Commonly the compiled code comes from a variety of other languages such as C, C++ and Fortran. One could also use this approach to wrap their compiled code with R or Python wrappers and run it on Azure ML.

To illustrate the process, we will build two MurmurHash modules from C++ for R and Python using the following two implementations on GitHub, and link them to Azure ML from a zipped folder

Link via David Smith.  I knew it was possible to call compiled C code from Python and R, but didn’t expect to be able to do it within Azure ML, so that’s good to know.

Comments closed

Machine Learning Packages In R

Published 2016-06-08 by Kevin Feasel

Khushbu Shah discusses good R packages to help with your machine learning projects:

If missing values are something which haunts you then MICE package is the real friend of yours.

When we face an issue of missing values we generally go ahead with basic imputations such as replacing with 0, replacing with mean, replacing with mode etc. but each of these methods are not versatile and could result into a possible data discrepancy.

MICE package helps you to impute missing values by using multiple techniques, depending on the kind of data you are working with.

I’d heard of a couple of these, but most of them are new to me.

Comments closed