Interview BigML.com

Here is an interview with Charlie Parker, head of large scale online algorithms at http://bigml.com

Ajay-  Describe your own personal background in scientific computing, and how you came to be involved with machine learning, cloud computing and BigML.com

Charlie- I am a machine learning Ph.D. from Oregon State University. Francisco Martin (our founder and CEO), Adam Ashenfelter (the lead developer on the tree algorithm), and myself were all studying machine learning at OSU around the same time. We all went our separate ways after that.

Francisco started Strands and turned it into a 100+ million dollar company building recommender systems. Adam worked for CleverSet, a probabilistic modeling company that was eventually sold to Cisco, I believe. I worked for several years in the research labs at Eastman Kodak on data mining, text analysis, and computer vision.

When Francisco left Strands to start BigML, he brought in Justin Donaldson who is a brilliant visualization guy from Indiana, and an ex-Googler named Jose Ortega who is responsible for most of our data infrastructure. They pulled in Adam and I a few months later. We also have Poul Petersen, a former Strands employee, who manages our herd of servers. He is a wizard and makes everyone else’s life much easier.

Ajay- You use clojure for the back end of BigML.com .Are there any other languages and packages you are considering? What makes clojure such a good fit for cloud computing ?

Charlie- Clojure is a great language because it offers you all of the benefits of Java (extensive libraries, cross-platform compatibility, easy integration with things like Hadoop, etc.) but has the syntactical elegance of a functional language. This makes our code base small and easy to read as well as powerful.

We’ve had occasional issues with speed, but that just means writing the occasional function or library in Java. As we build towards processing data at the Terabyte level, we’re hoping to create a framework that is language-agnostic to some extent. So if we have some great machine learning code in C, for example, we’ll use Clojure to tie everything together, but the code that does the heavy lifting will still be in C. For the API and Web layers, we use Python and Django, and Justin is a huge fan of HaXe for our visualizations.

 Ajay- Current support is for Decision Trees. When can we see SVM, K Means Clustering and Logit Regression?

Charlie- Right now we’re focused on perfecting our infrastructure and giving you new ways to put data in the system, but expect to see more algorithms appearing in the next few months. We want to make sure they are as beautiful and easy to use as the trees are. Without giving too much away, the first new thing we will probably introduce is an ensemble method of some sort (such as Boosting or Bagging). Clustering is a little further away but we’ll get there soon!

Ajay- How can we use the BigML.com API using R and Python.

Charlie- We have a public github repo for the language bindings. https://github.com/bigmlcom/io Right now, there there are only bash scripts but that should change very soon. The python bindings should be there in a matter of days, and the R bindings in probably a week or two. Clojure and Java bindings should follow shortly after that. We’ll have a blog post about it each time we release a new language binding. http://blog.bigml.com/

Ajay-  How can we predict large numbers of observations using a Model  that has been built and pruned (model scoring)?

Charlie- We are in the process of refactoring our backend right now for better support for batch prediction and model evaluation. This is something that is probably only a few weeks away. Keep your eye on our blog for updates!

Ajay-  How can we export models built in BigML.com for scoring data locally.

Charlie- This is as simple as a call to our API. https://bigml.com/developers/models The call gives you a JSON object representing the tree that is roughly equivalent to a PMML-style representation.

About-

You can read about Charlie Parker at http://www.linkedin.com/pub/charles-parker/11/85b/4b5 and the rest of the BigML team at

https://bigml.com/team

 

Software Review- BigML.com – Machine Learning meets the Cloud

I had a chance to dekko the new startup BigML https://bigml.com/ and was suitably impressed by the briefing and my own puttering around the site. Here is my review-

1) The website is very intutively designed- You can create a dataset from an uploaded file in one click and you can create a Decision Tree model in one click as well. I wish other cloud computing websites like  Google Prediction API make design so intutive and easy to understand. Also unlike Google Prediction API, the models are not black box models, but have a description which can be understood.

2) It includes some well known data sources for people trying it out. They were kind enough to offer 5 invite codes for readers of Decisionstats ( if you want to check it yourself, use the codes below the post, note they are one time only , so the first five get the invites.

BigML is still invite only but plan to get into open release soon.

3) Data Sources can only be by uploading files (csv) but they plan to change this hopefully to get data from buckets (s3? or Google?) and from URLs.

4) The one click operation to convert data source into a dataset shows a histogram (distribution) of individual variables.The back end is clojure , because the team explained it made the easiest sense and fit with Java. The good news (?) is you would never see the clojure code at the back end. You can read about it from http://clojure.org/

As cloud computing takes off (someday) I expect clojure popularity to take off as well.

Clojure is a dynamic programming language that targets the Java Virtual Machine (and the CLR, and JavaScript). It is designed to be a general-purpose language, combining the approachability and interactive development of a scripting language with an efficient and robust infrastructure for multithreaded programming. Clojure is a compiled language – it compiles directly to JVM bytecode, yet remains completely dynamic. Every feature supported by Clojure is supported at runtime. Clojure provides easy access to the Java frameworks, with optional type hints and type inference, to ensure that calls to Java can avoid reflection.

Clojure is a dialect of Lisp

 

5) As of now decision trees is the only distributed algol, but they expect to roll out other machine learning stuff soon. Hopefully this includes regression (as logit and linear) and k means clustering. The trees are created and pruned in real time which gives a slightly animated (and impressive effect). and yes model building is an one click operation.

The real time -live pruning is really impressive and I wonder why /how it can ever be replicated in other software based on desktop, because of the sheer interactive nature.

 

Making the model is just half the work. Creating predictions and scoring the model is what is really the money-earner. It is one click and customization is quite intuitive. It is not quite PMML compliant yet so I hope some Zemanta like functionality can be added so huge amounts of models can be applied to predictions or score data in real time.

 

If you are a developer/data hacker, you should check out this section too- it is quite impressive that the designers of BigML have planned for API access so early.

https://bigml.com/developers

BigML.io gives you:

  • Secure programmatic access to all your BigML resources.
  • Fully white-box access to your datasets and models.
  • Asynchronous creation of datasets and models.
  • Near real-time predictions.

 

Note: For your convenience, some of the snippets below include your real username and API key.

Please keep them secret.

REST API

BigML.io conforms to the design principles of Representational State Transfer (REST)BigML.io is enterely HTTP-based.

BigML.io gives you access to four basic resources: SourceDatasetModel and Prediction. You cancreatereadupdate, and delete resources using the respective standard HTTP methods: POSTGET,PUT and DELETE.

All communication with BigML.io is JSON formatted except for source creation. Source creation is handled with a HTTP PUT using the “multipart/form-data” content-type

HTTPS

All access to BigML.io must be performed over HTTPS

and https://bigml.com/developers/quick_start ( In think an R package which uses JSON ,RCurl  would further help in enhancing ease of usage).

 

Summary-

Overall a welcome addition to make software in the real of cloud computing and statistical computation/business analytics both easy to use and easy to deploy with fail safe mechanisms built in.

Check out https://bigml.com/ for yourself to see.

The invite codes are here -one time use only- first five get the invites- so click and try your luck, machine learning on the cloud.

If you dont get an invite (or it is already used, just leave your email there and wait a couple of days to get approval)

  1. https://bigml.com/accounts/register/?code=E1FE7
  2. https://bigml.com/accounts/register/?code=09991
  3. https://bigml.com/accounts/register/?code=5367D
  4. https://bigml.com/accounts/register/?code=76EEF
  5. https://bigml.com/accounts/register/?code=742FD

Using R for Cloud Computing – made very easy and free by BioConductor

I really liked the no hassles way Biocnoductor has put a cloud AMI loaded with RStudio to help people learn R, and even try using R from within a browser in the cloud.

Not only is the tutorial very easy to use- they also give away 2 hours for free computing!!!

Check it out-

Step 1

Step 2

Step 3

and wow! I am using Google Chrome to run R ..and its awesome!

Interesting- check out two hours for free — all you need is a browser and internet connection

http://www.bioconductor.org/help/cloud/

Web Analytics using R , Google Analytics and TS Forecasting

This is a continuation of the previous post on using Google Analytics .

Now that we have downloaded and plotted the data- we try and fit time series to the website data to forecast future traffic.

Some observations-

1) Google Analytics has 0 predictive analytics, it is just descriptive analytics and data visualization models (including the recent social analytics). However you can very well add in basic TS function using R to the GA API.

Why do people look at Website Analytics? To know today’s traffic and derive insights for the Future

2) Web Data clearly follows a 7 day peak and trough for weekly effects (weekdays and weekends), this is also true for hourly data …and this can be used for smoothing historic web data for future forecast.

3) On an advanced level, any hugely popular viral posts can be called a level shift (not drift) and accoringly dampened.

Test and Control!

Similarly using ARIMAX, we can factor in quantity and tag of posts as X regressor variables.

and now the code-( dont laugh at the simplicity please, I am just tinkering and playing with data here!)

You need to copy and paste the code at the bottom of   this post  http://www.decisionstats.com/using-google-analytics-with-r/ if you want to download your GA data down first.

Note I am using lubridate ,forecast and timeSeries packages in this section.

#Plotting the Traffic  plot(ga.data$data[,2],type="l") 

library(timeSeries)
library(forecast)

#Using package lubridate to convert character dates into time
library(lubridate)
ga.data$data[,1]=ymd(ga.data$data[,1])
ls()
dataset1=ga.data$data
names(dataset1) <- make.names(names(dataset1))
str(dataset1)
head(dataset1)
dataset2 <- ts(dataset1$ga.visitors,start=0,frequency = frequency(dataset1$ga.visitors), names=dataset1$ga.date)
str(dataset2)
head(dataset2)
ts.test=dataset2[1:200]
ts.control=dataset2[201:275]

 #Note I am splitting the data into test and control here

fitets=ets(ts.test)
plot(fitets)
testets=ets(ts.control,model=fitets)
accuracy(testets)
plot(testets)
spectrum(ts.test,method='ar')
decompose(ts.test)

library("TTR")
bb=SMA(dataset2,n=7)#We are doing a simple moving average for every 7 days. Note this can be 24 hrs for hourly data, or 30 days for daily data for month # 

to month comparison or 12 months for annual
#We notice that Web Analytics needs sommethening for every 7 thday as there is some relation to traffic on weekedays /weekends /same time last week
head(dataset2,40)
head(bb,40)

par(mfrow=c(2,1))
plot(bb,type="l",main="Using Seven Day Moving Average for Web Visitors")
plot(dataset2,main="Original Data")

Created by Pretty R at inside-R.org

Though I still wonder why the R query, gA R code /package could not be on the cloud (why it  needs to be downloaded)– cloud computing Gs?

Also how about adding some MORE predictive analytics to Google Analytics, chaps!

To be continued-

auto.arima() and forecasts!!!

cross validations!!!

and adapting the idiosyncratic periods and cycles  of web analytics to time series !!

Using Cloud Computing for Hacking

This is not about hacking the cloud. Instead this is about using the cloud to hack

 

Some articles last year wrote on how hackers used Amazon Ec2 for hacking/ddos attacks.

http://www.pcworld.com/businesscenter/article/216434/cloud_computing_used_to_hack_wireless_passwords.html

Roth claims that a typical wireless password can be guessed by EC2 and his software in about six minutes. He proved this by hacking networks in the area where he lives. The type of EC2 computers used in the attack costs 28 cents per minute, so $1.68 is all it could take to lay open a wireless network.

and

http://www.bloomberg.com/news/2011-05-15/sony-attack-shows-amazon-s-cloud-service-lures-hackers-at-pennies-an-hour.html

Cloud services are also attractive for hackers because the use of multiple servers can facilitate tasks such as cracking passwords, said Ray Valdes, an analyst at Gartner Inc. Amazon could improve measures to weed out bogus accounts, he said.

 

and this article by Anti-Sec pointed out how one can obtain a debit card anonymously

https://www.facebook.com/notes/lulzsec/want-to-be-a-ghost-on-the-internet/230293097062823

VPN Account without paper trail

  • Purchase prepaid visa card with cash
  • Purchase Bitcoins with Money Order
  • Donate Bitcoins to different account

 

Masking your IP address to log on is done by TOR

https://www.torproject.org/download/download.html.en

and the actual flooding is done by tools like LOIC or HOIC

http://sourceforge.net/projects/loic/

and

http://www.4shared.com/rar/UmCu0ds1/hoic.html

 

So what safeguards can be expected from the next wave of Teenage Mutant Ninjas..?

 

Cloud Computing – can be evil

Cloud Computing can be evil because-

1) Most browsers are owned by for profit corporations . Corporations can be evil, sometimes

And corporations can go bankrupt. You can back up data locally, but try backing up a corporation.

2) The content on your web page can be changed using translator extensions . This has interesting ramifications as in George Orwell. You may not be even aware of subtle changes introduced in your browser in the way it renders the html or some words using keywords from a browser extension app.

Imagine a new form of language called Politically Correct Truthspeak, and that can be in English but using machine learning learn to substitute politically sensitive words with Govt sanctioned words.

3) Your DNS and IP settings can be redirected using extensions. This means if a Govt passes a law- you can be denied the websites using just the browser not even the ISP.

Thats an extreme scenario for a authoritative govt creating its own version of Mafiaafire Redirector.

So how to keep the cloud computer honest?Move some stuff to the desktop

How to keep desktop computing efficient?Use some more cloud computing

It is not an OR but an AND function in which some computing can be local, some shared and some in the cloud.

Si?

Interview Michal Kosinski , Concerto Web Based App using #Rstats

Here is an interview with Michal Kosinski , leader of the team that has created Concerto – a web based application using R. What is Concerto? As per http://www.psychometrics.cam.ac.uk/page/300/concerto-testing-platform.htm

Concerto is a web based, adaptive testing platform for creating and running rich, dynamic tests. It combines the flexibility of HTML presentation with the computing power of the R language, and the safety and performance of the MySQL database. It’s totally free for commercial and academic use, and it’s open source

Ajay-  Describe your career in science from high school to this point. What are the various stats platforms you have trained on- and what do you think about their comparative advantages and disadvantages?  

Michal- I started with maths, but quickly realized that I prefer social sciences – thus after one year, I switched to a psychology major and obtained my MSc in Social Psychology with a specialization in Consumer Behaviour. At that time I was mostly using SPSS – as it was the only statistical package that was taught to students in my department. Also, it was not too bad for small samples and the rather basic analyses I was performing at that time.

 

My more recent research performed during my Mphil course in Psychometrics at Cambridge University followed by my current PhD project in social networks and research work at Microsoft Research, requires significantly more powerful tools. Initially, I tried to squeeze as much as possible from SPSS/PASW by mastering the syntax language. SPSS was all I knew, though I reached its limits pretty quickly and was forced to switch to R. It was a pretty dreary experience at the start, switching from an unwieldy but familiar environment into an unwelcoming command line interface, but I’ve quickly realized how empowering and convenient this tool was.

 

I believe that a course in R should be obligatory for all students that are likely to come close to any data analysis in their careers. It is really empowering – once you got the basics you have the potential to use virtually any method there is, and automate most tasks related to analysing and processing data. It is also free and open-source – so you can use it wherever you work. Finally, it enables you to quickly and seamlessly migrate to other powerful environments such as Matlab, C, or Python.

Ajay- What was the motivation behind building Concerto?

Michal- We deal with a lot of online projects at the Psychometrics Centre – one of them attracted more than 7 million unique participants. We needed a powerful tool that would allow researchers and practitioners to conveniently build and deliver online tests.

Also, our relationships with the website designers and software engineers that worked on developing our tests were rather difficult. We had trouble successfully explaining our needs, each little change was implemented with a delay and at significant cost. Not to mention the difficulties with embedding some more advanced methods (such as adaptive testing) in our tests.

So we created a tool allowing us, psychometricians, to easily develop psychometric tests from scratch an publish them online. And all this without having to hire software developers.

Ajay -Why did you choose R as the background for Concerto? What other languages and platforms did you consider. Apart from Concerto, how else do you utilize R in your center, department and University?

Michal- R was a natural choice as it is open-source, free, and nicely integrates with a server environment. Also, we believe that it is becoming a universal statistical and data processing language in science. We put increasing emphasis on teaching R to our students and we hope that it will replace SPSS/PASW as a default statistical tool for social scientists.

Ajay -What all can Concerto do besides a computer adaptive test?

Michal- We did not plan it initially, but Concerto turned out to be extremely flexible. In a nutshell, it is a web interface to R engine with a built-in MySQL database and easy-to-use developer panel. It can be installed on both Windows and Unix systems and used over the network or locally.

Effectively, it can be used to build any kind of web application that requires a powerful and quickly deployable statistical engine. For instance, I envision an easy to use website (that could look a bit like SPSS) allowing students to analyse their data using a web browser alone (learning the underlying R code simultaneously). Also, the authors of R libraries (or anyone else) could use Concerto to build user-friendly web interfaces to their methods.

Finally, Concerto can be conveniently used to build simple non-adaptive tests and questionnaires. It might seem to be slightly less intuitive at first than popular questionnaire services (such us my favourite Survey Monkey), but has virtually unlimited flexibility when it comes to item format, test flow, feedback options, etc. Also, it’s free.

Ajay- How do you see the cloud computing paradigm growing? Do you think browser based computation is here to stay?

Michal – I believe that cloud infrastructure is the future. Dynamically sharing computational and network resources between online service providers has a great competitive advantage over traditional strategies to deal with network infrastructure. I am sure the security concerns will be resolved soon, finishing the transformation of the network infrastructure as we know it. On the other hand, however, I do not see a reason why client-side (or browser) processing of the information should cease to exist – I rather think that the border between the cloud and personal or local computer will continually dissolve.

About

Michal Kosinski is Director of Operations for The Psychometrics Centre and Leader of the e-Psychometrics Unit. He is also a research advisor to the Online Services and Advertising group at the Microsoft Research Cambridge, and a visiting lecturer at the Department of Mathematics in the University of Namur, Belgium. You can read more about him at http://www.michalkosinski.com/

You can read more about Concerto at http://code.google.com/p/concerto-platform/ and http://www.psychometrics.cam.ac.uk/page/300/concerto-testing-platform.htm