Press "Enter" to skip to content

Category: Data Science

Check Those Feature Distributions

Antoine Rebecq shares a warning:

I was recently working on a cool dataset that looked unusually friendly. It was tidy, neat, interesting… the kind of things that you rarely encounter in the wild! My goal was to build a super simple predictor for one of the features. However, I kept getting poor results and at first couldn’t figure out what was happening.

There’s some good, practical advice in there, so check it out. H/T R-Bloggers

Leave a Comment

Geospatial Fraud Detection

Antoine Amend uses Databricks to identify financial fraud in a geographical area:

As part of this real-world solution, we are releasing a new open source geospatial library, GEOSCAN, to detect geospatial behaviors at massive scale, track customers patterns over time and detect anomalous card transactions. Finally, we demonstrate how organizations can surface anomalies from an analytics environment to an online data store (ODS) with tight SLA requirements following a Lambda-like infrastructure underpinned by Delta Lake, Apache Spark and MLflow.

Click through for the article, as well as three notebooks.

Leave a Comment

Simulating Prediction Intervals

Bryan Shalloway continues a series:

Part 1 of my series of posts on building prediction intervals used data held-out from model training to evaluate the characteristics of prediction intervals. In this post I will use hold-out data to estimate the width of the prediction intervals directly. Doing such can provide more reasonable and flexible intervals compared to analytic approaches.

Click through for the article, and be sure to check out part 1 if you haven’t already.

Leave a Comment

Working with Prediction Intervals

Bryan Shalloway explains how generating prediction intervals is different from making point predictions:

Before using the model for predictive inference, one should have reviewed overall performance on a holdout dataset to ensure the model is sufficiently accurate for the business context. For example, for our problem is an average error of ~12% and 90% prediction intervals of +/- ~25% of Sale_Price useful? If the answer is “no,” that suggests the need for more effort in improving the accuracy of the model (e.g. trying other transformations, features, model types). For our examples we are assuming the answer is ‘yes,’ our model is accurate enough (so it is appropriate to move-on and focus on prediction intervals).

Click through for the article.

Leave a Comment

Generating Random Numbers in R

Holger von Jouanne-Diedrich brings the noise:

In data science, we try to find, sometimes well-hidden, patterns (= signal) in often seemingly random data (= noise). Pseudo-Random Number Generators (PRNG) try to do the opposite: hiding a deterministic data generating process (= signal) by making it look like randomness (= noise). If you want to understand some basics behind the scenes of this fascinating topic, read on!

Click through for an explanation of the process.

Comments closed

k-gram Language Models in R

Valerio Gherardi takes us through the concept of k-grams:

The post is structured as follows: we start by giving a succinct theoretical introduction to kk-gram models. Subsequently, we illustrate how to train a kk-gram model in R using kgrams, and explain how to use the standard perplexity metric for model evaluation or tuning. Finally, we use our trained model to generate some random text at different temperatures.

This goes into some depth on the topic and is worth giving a careful read.

Comments closed

The Basics of k-Means Clustering

Nathaniel Schmucker explains some of the principles of k-means clustering:

k-Means is easy to implement. In R, you can use the function kmeans() to quickly deploy an efficient k-Means algorithm. On datasets of reasonable size (thousands of rows), the kmeans function runs in fractions of a second.

k-Means is easy to interpret (in 2 dimensions). If you have two features of your k-Means analysis (e.g., you are grouping by length and width), the result of the k-Means algorithm can be plotted on an xy-coordinate system to show the extent of each cluster. It’s easy to visually inspect the assignment to see if the k-Means analysis returned a meaningful insight. In more dimensions (e.g., length, width, and height) you will need to either create a 3D plot, summarize your features in a table, or find another alternative to describing your analysis. This loses the intuitive power that a 2D k-Means analysis has in convincing you or your audience that your analysis should be trusted. It’s not to say that your analysis is wrong; it simply takes more mental focus to understand what your analysis says.

The k-Means analysis, however, is not always the best choice. k-Means does well on data that naturally falls into spherical clusters. If your data has a different shape (linear, spiral, etc.), k-Means will force clustering into circles, which can result in outputs that defy human expectations. The algorithm is not wrong; we have fed the algorithm data it was never intended to understand.

There’s a lot of depth in this article which makes it really interesting.

Comments closed

Bayesian vs Frequentist Approaches to Machine Learning

Ajit Jaokar has an interesting series. Here’s part one:

The arguments / discussions between the Bayesian vs frequentist approaches in statistics are long running. I am interested in how these approaches impact machine learning. Often, books on machine learning combine the two approaches, or in some cases, take only one approach. This does not help from a learning standpoint.  

So, in this two-part blog we first discuss the differences between the Frequentist and Bayesian approaches. Then, we discuss how they apply to machine learning algorithms.

Part two extends from there:

Sampled from a distribution: Many machine learning algorithms make assumptions that the data is sampled from a frequency. For example, linear regression assumes gaussian distribution and logistic regression assumes that the data is sampled from a Bernoulli distribution. Hence, these algorithms take a frequentist approach

My biases push me toward Bayesian approaches, and I really like what I see in Stan, but these techniques do often require a lot more processing power.

1 Comment

Analyzing XGBoost Training Reports

Simon Zamarin, et al, walk us through using XGBoost reports in Amazon’s Sagemaker Debugger:

In 2019, AWS unveiled Amazon SageMaker Debugger, a SageMaker capability that enables you to automatically detect a variety of issues that may arise while a model is being trained. SageMaker Debugger captures model state data at specified intervals during a training job. With this data, SageMaker Debugger can detect training issues or anomalies by leveraging built-in or user-defined rules. In addition to detecting issues during the training job, you can analyze the captured state data afterwards to evaluate model performance and identify areas for improvement. This task is made easier with the newly launched XGBoost training report feature. With a minimal amount of code changes, SageMaker Debugger generates a comprehensive report outlining key information that you can use to evaluate and improve the model.

This post shows you an end-to-end example of training an XGBoost model on Sagemaker and how to enable the automatic XGBoost report functionality in Sagemaker Debugger to quickly and easily evaluate model performance and identify areas of improvement for your model. Even if you don’t have a lot of data science experience, you can still gauge how well the model performs and identify areas of improvement based on information provided by the report. The code from this post is available in the GitHub repo.

Click through for an example of this in action.

Comments closed

Applied ML Prototypes

Alex Bleakley and Santiago Giraldo announce Applied ML Prototypes:

To directly address these challenges, we’ve released Applied ML Prototypes (AMPs) — a revolutionary new way of developing and shipping enterprise ML use cases — which provide complete ML projects that can be deployed with one click directly from Cloudera Machine Learning. AMPs enable data scientists to go from an idea to a fully working ML use case in a fraction of the time, with an end-to-end framework for building, deploying, and monitoring business-ready ML applications instantly. 

AMPs move the starting line for any ML project by enabling data scientists to start with a full end-to-end project developed for a similar use case, including a trained and deployed ML model, as well as prebuilt predictive business applications, out of the box. This means that ML development teams can tackle their own ML business use cases more quickly, from those involving churn modeling, to sentiment analysis, to anomaly detection and beyond.

Getting past the marketing fluff, there are some interesting ideas here.

Comments closed