Category: Data Science

Measuring Model Accuracy

Published 2017-07-11 by Kevin Feasel

Fabio Veronesi shows several methods of testing model accuracy:

Mean Squared Deviation or Mean Squared Error

This is simply the numerator of the previous equation, but it is not used often. The issue with both the RMSE and the MSE is that since they square the residuals they tend to be more affected by large residuals. This means that even if our model explains the large majority of the variation in the data very well, with few exceptions; these exceptions will inflate the value of RMSE.

Click through for several calculations. H/T R-bloggers

Comments closed

Regularization Prevents Overfitting

Published 2017-07-07 by Kevin Feasel

Hui Li has an explanation of what regularization is and how it works to reduce the likelihood of overfitting training data:

Assume that the red line is the regression model we learn from the training data set. It can be seen that the learned model fits the training data set perfectly, while it cannot generalize well to the data not included in the training set. There are several ways to avoid the problem of overfitting.

To remedy this problem, we could:

Get more training examples.

Use a simple predictor.

Select a subsample of features.

In this blog post, we focus on the second and third ways to avoid overfitting by introducing regularization on the parameters $β_{i}$ of the model.

Read the whole thing.

Comments closed

Loan Chargeoff Templates

Published 2017-07-05 by Kevin Feasel

Ajay Jagannathan announces a couple new Cortana Intelligence Solutions Gallery templates:

For more information, read this blog: End to End Loan ChargeOff Prediction Built Using Azure HDInsight Spark Clusters and SQL Server 2016 R Service

We have published two solution templates deployable using two technology stacks for the above chargeoff scenario:-

Loan Chargeoff Prediction using SQL Server 2016 R Services – Using DSVM with SQL Server 2016 and Microsoft ML, this solution template walks through how to create and clean up a set of simulated data, use 5 different models to train, select the best performant model, perform scoring using the model and save the prediction results back to SQL Server. A PowerBI report connects to the prediction table and show interactive reports with the user on the chargeoff prediction.
Loan Chargeoff Prediction using HDInsight Spark Clusters – This solution demonstrates how to develop machine learning models for predicting loan chargeoff (including data processing, feature engineering, training and evaluating models), deploy the models as a web service (on the edge node) and consume the web service remotely with Microsoft R Server on Azure HDInsight Spark clusters. The final predictions is saved to a Hive table which could be visualized in Power BI.

These tend to be nice because they show you how the different pieces of the Azure stack tie together.

Comments closed

What’s Wrong With CRISP-DM

Published 2017-07-03 by Kevin Feasel

Jen Stirrup explains the issues with CRISP-DM, a model for data mining:

The model no longer seems to be actively maintained. At the time of writing, the official site, CRISP-DM.org, is no longer being maintained. Further, the framework itself has not been updated on issues on working with new technologies, such as Big Data.

Jen then contrasts with with Microsoft’s Team Data Science Process framework; click through for that.

Comments closed

Mann-Whitney U Test in SQL

Published 2017-07-03 by Kevin Feasel

Phil Factor continues his Statistics in SQL series with the Mann-Whitney U test:

There are several ways that you can test this, but nobody is going to argue with you if you use a Mann–Whitney U test to test whether two samples come from the same distribution. It doesn’t require that the data has any particular type of distribution. It just requires that each observation is done by a different member of the population so that all the observations from both groups are independent of each other. It is really just a test of differences in mean-rank between two populations’ pooled ranking. To test this difference It has to be possible to compare any of the observations with any of the others and say which of the two are greater. Your objective is to disprove the assumption that The distributions of both populations are equal. Calculating a measure of the difference is simple, and was designed to be done easily by hand before computers. The probability that the observed difference occurred by chance is easily calculated for large samples because U then approximates to the normal distribution, but it is complex for small samples. Here, we have a small sample and are just interested in whether the two-tailed test is signifcant at the five percent level so we dodge the bullet by using a significance lookup table for the critical value of U.

Read on for Phil’s implementation of the test.

Comments closed

Neural Nets On Spark

Published 2017-06-28 by Kevin Feasel

Nisha Muktewar and Seth Hendrickson show how to use Deeplearning4j to build deep learning models on Hadoop and Spark:

Modern convolutional networks can have several hundred million parameters. One of the top-performing neural networks in the Large Scale Visual Recognition Challenge (also known as “ImageNet”), has 140 million parameters to train! These networks not only take a lot of compute and storage resources (even with a cluster of GPUs, they can take weeks to train), but also require a lot of data. With only 30000 images, it is not practical to train such a complex model on Caltech-256 as there are not enough examples to adequately learn so many parameters. Instead, it is better to employ a method called transfer learning, which involves taking a pre-trained model and repurposing it for other use cases. Transfer learning can also greatly reduce the computational burden and remove the need for large swaths of specialized compute resources like GPUs.

It is possible to repurpose these models because convolutional neural networks tend to learn very general features when trained on image datasets, and this type of feature learning is often useful on other image datasets. For example, a network trained on ImageNet is likely to have learned how to recognize shapes, facial features, patterns, text, and so on, which will no doubt be useful for the Caltech-256 dataset.

This is a longer post, but on an extremely interesting topic.

Comments closed

Linear Prediction Confidence Region Flare-Out

Published 2017-06-26 by Kevin Feasel

John Cook explains why the confidence region of a tracked object flares out instead of looking conical (or some other shape):

Suppose you’re tracking some object based on its initial position x₀ and initial velocity v₀. The initial position and initial velocity are estimated from normal distributions with standard deviations σ_x and σ_v. (To keep things simple, let’s assume our object is moving in only one dimension and that the distributions around initial position and velocity are independent.)

The confidence region for the object flares out over time, something like the bell of a trumpet.

Read on for the explanation.

Comments closed

Bayesian Average

Published 2017-06-21 by Kevin Feasel

Jelte Hoekstra has a fun post applying the Bayesian average to board game ratings:

Maybe you want to explore the best boardgames but instead you find the top 100 filled with 10/10 scores. Experience many such false positives and you will lose faith in the rating system. Let’s be clear this isn’t exactly incidental either: most games have relatively few votes and suffer from this phenomenon.

The Bayesian average

Fortunately, there are ways to deal with this. BoardGameGeek’s solution is to replace the average by the Bayesian average. In Bayesian statistics we start out with a prior that represents our a priori assumptions. When evidence comes in we can update this prior, computing a so called posterior that reflects our updated belief.

Applied to boardgames this means: if we have an unrated game we might as well assume it’s average. If not, the ratings will have to convince us otherwise. This certainly removes outliers as we will see below!

This is a rather interesting article and you can easily apply it to other rating systems as well.

Comments closed

Calculating Relative Risk In T-SQL

Published 2017-06-20 by Kevin Feasel

Mala Mahadevan explains how to calculate relative risk using T-SQL:

In this post we will explore a common statistical term – Relative Risk, otherwise called Risk Factor. Relative Risk is a term that is important to understand when you are doing comparative studies of two groups that are different in some specific way. The most common usage of this is in drug testing – with one group that has been exposed to medication and one group that has not. Or , in comparison of two different medications with two groups with each exposed to a different one.

Read on for an example of a statistical formula calculation which might actually be easier in T-SQL than R.

Comments closed

Random Forests In scikit-learn

Published 2017-06-20 by Kevin Feasel

Mark Needham shows how easy it is to create a random forest model in Python using scikit-learn:

As I mentioned in a blog post a couple of weeks ago, I’ve been playing around with the Kaggle House Prices competition and the most recent thing I tried was training a random forest regressor.

Unfortunately, although it gave me better results locally it got a worse score on the unseen data, which I figured meant I’d overfitted the model.

I wasn’t really sure how to work out if that theory was true or not, but by chance, I was reading Chris Albon’s blog and found a post where he explains how to inspect the importance of every feature in a random forest. Just what I needed!

There’s a nagging voice in my head saying “Principal Component Analysis” as I read this post.

Comments closed