Revolutions

A reproducibility horror story (and the heroes, knitr and checkpoint)

You download the data and complete your analysis with ample time to spare. Then, just before deadline, your collaborator lets you know that they've "fixed a data error". Now, you have to do your analysis all over again. This is the reproducibility horror story:

 

R provides many tools to make reproducibility easy, and the creators of the above video, Ecoinformática - AEET, provide a useful list of tutorials and guides. Chief amongst these is using the knitr package for R: the R language automates the process of importing, preparing and analyzing the data, while knitr automates the process of assembling text, code, tables and charts into a Word, PDF, HTML and many other document formats.  

But while knitr solves a good chunk[*] of the reproducibility problem, there's one complicating factor it doesn't deal with: updated R pacakges. In the same way that a collaborator updating the data triggers a restart, someone updating an R package your script uses can also affect your results. (That someone was likely you, working on a different R project.) The checkpoint package for R solves that problem by letting you "lock in" the package versions you use with a project. It's easy to use: all you need to do is add a line like checkpoint("2016-08-31") to the beginning of your script, which:

• Downloads all the packages used by your project (those mentioned files in the current folder), as they were on August 31, 2016
• Installs them in a folder specific to this project (so they're independent from other R projects), and
• Makes sure R uses those package versions when you run your script

It does some clever things to avoid re-downloading packages if it doesn't need to, and avoiding duplicates of multiple copies of the same package version, but that's the basic gist. Checkpoint also makes it really easy to share code with others, because you can be confident they'll also get the packages they need to make your script work. You can learn more about the checkpoint package here and in this vignette, and just install it from CRAN to get started. (If you use Microsoft R Open you don't even need to download it, it's already included.)

[*] pun intended

Telco Customer Churn with R in SQL Server 2016

by Fang Zhou, Data Scientist at Microsoft and Wee Hyong Tok, Senior Data Scientist Manager at Microsoft

In today’s fast-paced world, mobile phone customers have many choices and can easily switch between service providers. Improving customer attrition rates and enhancing a customer’s experience are valuable ways to reduce customer acquisition costs and maintain a high-quality service.

Many industries, including mobile phone service providers, use churn models to predict which customers are most likely to leave, and to understand which factors cause customers to stop using their service. In this blog post, we sketch a solution to help providers, especially telecommunication companies, predict customer churn. This solution provides guidance on how to harness advanced analytics in SQL Server 2016, leveraging its new feature R Services and familiar SQL tools and technologies. To be specific, it covers the following phases:

• How a data scientist can make use of SQL Server R Services to do data science, and
• How to operationalize the telco churn models via T-SQL stored procedures and build an intelligent application.

The usage and comparison of CRAN R and RevoScaleR packages are also highlighted here.

Data Science with R Services

In the case of telco customer churn, we collected a combination of the call detail record data and customer profile data from a mobile carrier, and then followed the data science process --  data exploration and visualization, data pre-processing and feature engineering, model training, scoring and evaluation -- in order to achieve the churn prediction. With a churn indicator in the dataset taking value 1 when the customer is churned and taking value 0 when the customer is non-churned, we addressed the problem as a binary classification problem and tried various tree-based models along with methods like bagging, random forests and boosting. Because the number of churned customers is much less than that of non-churned customers (making the data set quite unbalanced), SMOTE (Synthetic Minority Oversampling Technique) was applied to adjust the proportion of majority class over minority class in the training data set, thus further improving model performance, especially precision and recall.

All the above data science procedures could be implemented with base R. Rather than moving the data out from the database to an external machine running R, we instead run R scripts directly on SQL Server data by leveraging the in-database analytics capability provided by SQL Server R Services, taking advantage of the rich and powerful CRAN R packages plus the parallel external memory algorithms in the RevoScaleR library. In what follows, we will describe the specific R packages and algorithms that we used to implement the data science solution for predicting telco customer churn.

In order to explore and visualize the data, we could leverage a branch of CRAN R packages for manipulation and visualization, such as dplyr, ggplot and d3heatmap. To avoid the memory limitations of CRAN R and avoid moving the raw data out from the database, we instead used RevoScaleR functions for data aggregation. The rxCube and rxCrossTabs functions, used under the in-sql compute context, is all that's needed fir the visualization work. Below is a sample interactive visualization dashboard created by adding the power of leaflet and shiny.

Video series: Introduction to Microsoft R Server

Microsoft R Server extends the base R language and Microsoft R Open with big-data capabilities. Specifically, it adds the RevoScaleR package, which creates an out-of-memory "CDF" data structure (so you can process data larger than available RAM), and algorithms that allow you to perform computations on such data using parallel and distributed algorithms. (A limited version of the RevoScaleR package is also included in the free Microsoft R Client.)

If you'd like to get up to speed on the capabilities of Microsoft R Server, my colleague Matt Parker has created a 4-part video series that introduces Microsoft R Server and delves into data ingestion, data management, and predictive modeling. The videos are embedded below, and you can also download the videos for offline viewing from Channel 9.

Part 1: Introduction to Microsoft R Server

 

Part 2: Data Ingestion

Because it's Friday: The font of Stranger Things

If you haven't caught up on Neflix's Stranger Things yet, you're missing out. It's awesome: a perfect homage to 80's-era sci-fi and fantasy movies, with a great cast and stellar performances (particularly from Winona Ryder and newcomer Millie Bobby Brown). This spoiler-laden supercut shows scenes from Stranger Things alongside corresponding scenes from Close Encounters, the Goonies, Alien, and many other films of the era and the inspirations are clear. If you liked those movies, you'll love Stranger Things.

One thing I love about the series is the opening titles, which is simply a close-up of the letters in the two-word title slowly moving to assemble the logo, Stranger Things. It's done in the font ITC Benguiat, which brings back memories of it being used on the cover of Choose Your Own Adenture books and Stephen King novels. The title sequence looks like it was filmed, not animated, and in a way it was: filmed backlit transparencies were used as the reference for the animation process. The short film above (via Vox) is a lovely insight into the process behind the creation of the title sequence.

That's all from us for this week. Have a good weekend (you have enough time to catch up on Stranger Things!), and we'll be back here on Monday. See you then.

Microsoft R Open 3.3.1 now available for Windows, Mac and Linux

Microsoft R Open 3.3.1, our enhanced disstribution of open source R, is now available for download for Windows, Mac, and Linux. This update upgrades the R langauge engine to version 3.3.1, streamlines the installation process, and bundles some additional packages for parallel programming.

R version 3.3.1 fixes a few rarely-encountered bugs, for example to generate Gamma random numbers with zero or infinite rate parameters, and correctly match text that only differed in the encoding. (See here for a complete list of fixes.) There are no user-visible changes in the language itself, which means that scripts and packages should work without changes from MRO 3.3.0.

For reproducibility, installing packages with MRO 3.3.1 will by default get you package versions as of July 1, 2016. Many packages have been updated or released since MRO 3.3.0, including packages for multivariate covariance generalized linear models, simulating data sets, and creating HTML tables. (See here for some highlights of new packages.) As always, if you want to use newer versions of CRAN packages (or access older versions, for reproducibility), just use the checkpoint function to access the CRAN Time Machine. 

MRO 3.3.1 now pre-installs some additional packages, including the parallel-programming packages doParallel, foreach and iterators and (on Windows) the RODBC package for interacting with databases. (See here for a list of packages bundled with MRO.) MRO 3.3.1 is also now easier to install: it's now just a single download, with the option to install the Math Kernel Libraries (and speed up computations) during the install process. (Some benchmarks from an older version of MRO are shown below; results are about the same for MRO 3.3.1.)

We hope you find Microsoft R Open useful, and if you have any comments or questions please visit the Microsoft R Open forum. To download Microsoft R Open (it's free!), simply follow the link below.

MRAN: Download Microsoft R Open

R with Power BI: Import, Transform, Visualize and Share

Power BI, Microsoft's data visualization and reporting platform, has made great strides in the past year integrating the R language. This Computerworld article describes the recent advances with Power BI and R. In short, you can:

• import data into Power BI by using an R script
• cleanse and transform other data sources coming into Power BI using R functions
• create custom charts in a Power BI dashboard using the R language, like these maps
• share R scripts with others for use with Power BI in the R Scripts Showcase
• create dashboards with Power BI desktop and R on your local machine, and share them with others using Power BI Online

Power BI desktop is completely free to download and use, and includes all the features you need to create visualizations, reports and dashboards. (Publishing to Power BI online requires a subscription, though.) Power BI desktop and R are both included in the Data Science Virtual Machine, so that's another easy way to get started.

Sharon Laivand from the Power BI engineering team recently gave a webinar showcasing the capabilities of Power BI and R. Fast-forward to the 29-minute mark to see how to create a report incorporating R-based calculations and graphics, and then share it with others (even people who don't have R installed!) using Power BI online. 

For more information about Power BI, visit powerbi.microsoft.com.

Deep Learning Part 2: Transfer Learning and Fine-tuning Deep Convolutional Neural Networks

by Anusua Trivedi, Microsoft Data Scientist

This is a blog series in several parts — where I describe my experiences and go deep into the reasons behind my choices. In Part 1, I discussed the pros and cons of different symbolic frameworks, and my reasons for choosing Theano (with Lasagne) as my platform of choice.

Part 2 of this blog series is based on my upcoming talk at The Data Science Conference, 2016. Here in Part 2, I describe Deep Convolutional Neural Networks (DCNNs) and how Transfer learning and Fine-tuning helps better the training process for domain specific images.

Please feel free to email me at [email protected] if you have questions.

Introduction

The eye disease Diabetic Retinopathy (DR) is a common cause of vision loss. Screening diabetic patients using fluorescein angiography images can potentially reduce the risk of blindness. Current trends in the research have demonstrated that DCNNs are very effective in automatically analyzing large collections of images and identifying features that can categorize images with minimum error. DCNNs are rarely trained from scratch, as it is relatively uncommon to have a domain-specific dataset of sufficient size. Since modern DCNNs take 2-3 weeks to train across GPUs, Berkley Vision and Learning Center (BVLC) have released some final DCNN checkpoints. In this blog, we use such a pre-trained network: GoogLeNet. This GoogLeNet network is pre-trained on a large collection of natural ImageNet images. We transfer the learned ImageNet weights as initial weights for the network, and fine-tune these pre-trained generic network to recognize fluorescein angiography images of eyes and improve DR prediction.

Using explicit feature extraction to predict Diabetic Retinopathy

Much work has been done in developing algorithms and morphological image processing techniques that explicitly extract features prevalent in patients with DR. The generic workflow used in a standard image classification technique is as follows:

• Image preprocessing techniques for noise removal and contrast enhancement
• Feature extraction technique
• Classification
• Prediction

Faust et al. provide a very comprehensive analysis of models that use explicit feature extraction for DR screening. Vujosevic et al. build a binary classifier on a dataset of 55 patients by explicitly forming single lesion features. These authors use morphological image processing techniques to extract blood vessel, and hemorrhage features and then train an SVM on a data set of 331 images. These authors report accuracy of 90% and sensitivity of 90% on binary classification task with a dataset of 140 images.

However, all these processes are very time and effort consuming. Further improvements in prediction accuracy require large quantities of labeled data. Image processing and feature extraction of image datasets is very complex and time-consuming. Thus, we choose to automate the image processing and feature extraction step by using DCNNs.

Deep convolutional neural network (DCNN)

Image data requires subject-matter expertise to extract key features. DCNNs extract features automatically from domain-specific images, without any feature engineering techniques. This process makes DCNNs suitable for image analysis:

• DCNNs train networks with many layers
• Multiple layers work to build an improved feature space
• Initial layers learn 1st order features (e.g. color, edges etc.)
• Later layers learn higher order features (specific to input dataset)
• Lastly, final layer features are fed into classification layer(s)

 

C layers are convolutions, S layers are pool/sample

 

 

Convolution: Convolution layers consist of a rectangular grid of neurons. The weights for this are the same for each neuron in the convolution layer. The convolution layer weights specify the convolution filter.

 

Convolution

 

 

Pooling: The pooling layer takes small rectangular blocks from the convolutional layer and subsamples it to produce a single output from that block.

 

Pooling

 

 

In this post, we are using GoogLeNet DCNN, which was developed at Google. GoogLeNet won the ImageNet challenge in 2014, setting the record for the best contemporaneous results. Motivations for this model were a simultaneously deeper as well as computationally inexpensive architecture.

Edward Tufte Keynote Presenter at Data Science Summit, Sep 26-27

I'm excited to share that one of my data science heroes will be a presenter at the Microsoft Data Science Summit in Atlanta, September 26-27. Edward Tufte, the data visualization pioneer, will deliver a keynote address on the future of data analysis and the how to make more credible conclusions based on data.

If you're not familiar with Tufte, a great place to start is to read his seminal book Visual Display of Quantitative Information. First published in 1983 — well before the advent of mainstream data visualization software — this is the book that introduced and/or popularized many familiar concepts in data visualization today, such a small multiples, sparklines, and the data-ink ratio. Check out this 2011 Washington Monthy profile for more background on Tufte's career and influence. Tufte's work also influenced R: you can easily recreate many of Tufte's graphics in the R graphics system, including this famous weather chart.

The program for the Data Science Summit looks fantastic, and will also include keynote presentations from Microsoft CEO Satya Nadella and Data Group CVP Joseph Sirosh. Also there's a fantastic crop of Microsoft data scientists (plus yours truly) giving a wealth of practical presentations on how to use Microsoft tools and open-source software for data science. Here's just a sample:

• Jennifer Marsman will speak about building intelligent applications with the Cognitive Services APIs
• Danielle Dean will describe deploying real-world predictive maintenance solutions based on sensor data
• Brandon Rohrer will give a live presentation of his Data Science for Absolutely Everybody series
• Frank Seide will introduce CNTK, Microsoft's open source deep learning toolkit
• Maxim Likuyanov will share some best practices for interactive data analysis and scalable machine learning with Apache Spark
• Rafal Lukawiecki will explain how to apply data science in a business context 
• Debraj GuhaThakurta and Max Kaznady will demonstrate statistical modeling on huge data sets with Microsoft R Server and Spark
• David Smith (that's me!) will give some examples of how data science at Microsoft (and R!) is being used to improve the lives of disabled people
• ... and many many more!

Check out the agenda for the breakout sessions on the Data Science Summit page for more. I hope to see you there: it will be a great opportunity to meet with Microsoft's data science team and see some great talks as well. To register, follow the link below.

Microsoft Data Science Summit, September 26-27, Altanta GA: Register Now

Five great charts in 5 lines of R code each

Sharon Machlis is a journalist with Computerworld, and to show other journalists how great R is for data visualization she shows them these five data visualizations, each of which can be created in 5 lines of R code or less.

 

I've reproduced Sharon's code and charts below. I did make a couple of tweaks to the code, though. I added a call to checkpoint("2016-08-22") which, if you've saved the code to a file, will install all the necessary packages for you. (I also verified that the code runs with package versions as of today's date, and if you're trying out this code at a later time it will continue to do so, thanks to checkpoint.) I also modified the data download code to make it work more easily on Windows. Here are the charts and code:

Because it's Friday: Gravity wave detecting lasers ... in space!

Earlier this week I was lucky enough to meet and see a fascinating talk by Shane L. Larson, astronomer at Chicago's Adler Planetarium. The topic was Black Holes and Gravitational Waves, and the first part of the talk focused on the LIGO project, a pair of massive detectors in Washington and Louisiana that can detect the tiny ripples in space caused by massive gravitational events far in the universe, like stars going supernova or black holes colliding. And on September 14 last year, not long after the detectors were first turned on, that's exactly what they detected: two massive black holes, merging in space.

Here's what that event would look like, if you could position yourself near the merging black holes. This computer simulation was generated using the same software used for the movie Interstellar.)

 

However, something I learned about LIGO was that while telescopes are more like eyes on the universe, these gravity-wave detectors are more like ears: they can 'hear' the 'sound' from these cataclysmic events as the gravity waves pass through earth, but don't form an image of them directly. We can get a rough idea of the direction of the event because we have two sensors (in much the same way our two ears give us a rough idea of the direction of a loud sound). You can actually convert the sensor readings into audible sound, and this is what the above event sounded like: 

 

The actual sound is just 1 second long; it's played 4 times here, first in its actual frequency, then upsampled to a higher register where it's easier to hear (and then the cycle repeats). The little "chirp" you hear is the final phase of the black hole merger, where they spin super-rapidly before colliding at half the speed of light.

We're in the early days of gravitational wave observation, which literally gives us a new view (or listen?) on the universe to observe events that don't generate light or are obscured by the dust of the galaxy. More sensors are being constructed around the globe, which will give us the ability to detect more events and get a better sense of where they occurred. And there's even a project to build an observatory in space: LISA, the Laser Interferometry Space Antenna. (The original NASA project was sadly not funded, but the European Space Agency has taken up the mantle.) Like the ground-based detectors, LISA uses lasers to detect the minute (smaller than the diameter of a proton!) change in lengths of the detection arms as the gravity waves roll through. A precursor mission, LISA Pathfinder, was launched in December as a proof of concept to make sure accurate space-based detection of distances was possible. The experiment was to float two small masses in space, and use a laser reflecting off the sides to measure the distance. Dr Larson gave me a replica of one of the masses to hold — even though it was made of tungsten and not gold/platinum like the real ones, it was still surprisingly dense!

Thankfully, the LISA Pathfinder mission was a success, and the space-based gravity wave observational platform could be launched as early as 2039.

That's all from us on the blog for this week. See you on Monday, and have a great weekend!