A Fine Slice Of SQL Server
Sebastian Sauer takes us through confidence intervals:
Some one claims:
“A CI is the interval where 95% of all possible sample means fall into”.
Let’s check that.
Read the whole thing. Confidence intervals have their value, but this is one case where the frequentist model of statistics can trip you up between hypothetical and real, changing an accurate statement into a myth.
Comments closedTori Tompkins takes us through principles of anomaly detection in Apache Spark 3.0:
To calculate Median Absolute Deviation (MAD) you need to calculate the difference between the value and the median. In simpler terms, you will need to calculate the median of the entire dataset, the difference between each value and this median, then take another median of all the differences.
In Spark you can use a SQL expression ‘percentile()’ to calculate any medians or quartiles in a dataframe. ‘percentile()’ expects a column and an array of percentiles to calculate (for median we can provide ‘array(0.5)’ because we want the 50% value ie median) and will return an array of results.
Like standard deviation, to use MAD to identify the outliers it needs to be a certain number of MAD’s away. This number is also referred to as the threshold and is defaulted to 3.
Read on for three measures and their implementations in PySpark.
Comments closedCharles Thibault explains why you might not want to use logistic regression at all:
I wrote a post on my LinkedIn about why you should NEVER run a Logistic Regression. (Unless you really have to).
Click through for the kernel of the argument or check out the video.
Comments closedBryan Shalloway shows off some new functionality in the rsample package:
For some problems you may want to take a traditional regression or classification based approach while still accounting for the date/time-sensitive components of your data. In this post I will use the
tidymodelssuite of packages to:– build lag based and non-lag based features
– set-up appropriate time series cross-validation windows
– evaluate performance of linear regression and random forest models on a regression problemFor my example I will use data from Wake County food inspections. I will try to predict the
SCOREfor upcoming restaurant food inspections.
Click through to see it in action.
Comments closedLayla Yang and Hector Leano walk us through measuring how effective an advertising campaign was:
At a high level we are connecting a time series of regional sales to regional offline and online ad impressions over the trailing thirty days. By using ML to compare the different kinds of measurements (TV impressions or GRPs versus digital banner clicks versus social likes) across all regions, we then correlate the type of engagement to incremental regional sales in order to build attribution and forecasting models. The challenge comes in merging advertising KPIs such as impressions, clicks, and page views from different data sources with different schemas (e.g., one source might use day parts to measure impressions while another uses exact time and date; location might be by zip code in one source and by metropolitan area in another).
As an example, we are using a SafeGraph rich dataset for foot traffic data to restaurants from the same chain. While we are using mocked offline store visits for this example, you can just as easily plug in offline and online sales data provided you have region and date included in your sales data. We will read in different locations’ in-store visit data, explore the data in PySpark and Spark SQL, and make the data clean, reliable and analytics ready for the ML task. For this example, the marketing team wants to find out which of the online media channels is the most effective channel to drive in-store visits.A
Click through for the article as well as notebooks.
Comments closedLin Lee Cheong, et al, relay some interesting research:
NFL’s Next Gen Stats (NGS) powered by AWS accurately captures player and ball data in real time for every play and every NFL game—over 300 million data points per season—through the extensive use of sensors in players’ pads and the ball. With this rich set of tracking data, NGS uses AWS machine learning (ML) technology to uncover deeper insights and develop a better understanding of various aspects and trends of the game. To date, NGS metrics have focused on helping fans better appreciate and understand the offense and defense in gameplay through the application of advanced analytics, particularly in the passing game. Thanks to tracking data, it’s possible to quantify the difficulty of passes, model expected yards after catch, and determine the value of various play outcomes. A logical next step with this analytical information is to evaluate quarterback decision-making, such as whether the quarterback has considered all eligible receivers and evaluated tradeoffs accurately.
To effectively model quarterback decision-making, we considered a few key metrics—mainly the probability of different events occurring on a pass, and the value of said events. A pass can result in three outcomes: completion, incompletion, or interception. NGS has already created models that provide probabilities of these outcomes, but these events rely on information that’s available at only two points during the play: when the ball is thrown (termed as pass-forward), and when the ball arrives to a receiver (pass-arrived). Because of this, creating accurate probabilities requires modeling the trajectory of players between those two points in time.
For these probabilities, the quarterback’s decision is heavily influenced by the quality of defensive coverage on various receivers, because a receiver with a closely covered defender has a lower likelihood of pass completion compared to a receiver who is wide open due to blown coverage. Furthermore, defenders are inherently reactive to how the play progresses. Defenses move in completely different ways depending on which receiver is targeted on the pass. This means that a trajectory model for defenders has to similarly be reactive to the specified targeted receiver in a believable manner.
Click through for details on the study.
Comments closedStephanie Glen gives us examples of where specific probability distributions appear naturally:
If you’re in the beginning stages of your data science credential journey, you’re either about to take (or have taken) a probability class. As part of that class, you’re introduced to several different probability distributions, like the binomial distribution, geometric distribution and uniform distribution. You might be tempted to skip over some elementary topics and just scrape by with a bare pass. Because, let’s face it–the way probability is taught (with dice rolls and cards) is far removed from the glamor of data science. You may be wondering
When am I ever going to calculate the probability of five die rolls in a row in real life?
Click through for the answer and for a chart provides different scenarios for real-world probability distributions.
Comments closedRyohei Fujimaki summarizes some of the reasons why data science projects can fail:
According to Gartner analyst Nick Heudecker, over 85% of data science projects fail. A report from Dimensional Research indicated that only 4% of companies have succeeded in deploying ML models to production environment.
Even more critical, the economic downturn caused by the COVID-19 pandemic has placed increased pressure on data science and BI teams to deliver more with less. In this down market, organizations are reassessing which AI/ML models they should develop, how to optimize resources and how to best use valuable budget dollars for maximum impact. In this type of environment, AI/ML project failure is simply not acceptable.
That 85% sounds suspiciously like the percentage of failed business intelligence and data warehouse projects, as well as the percentage of failed big data projects. It’s close enough that it makes me want to come up with some overarching idea that projects based on the consolidation of multiple independent data systems across several business units are liable to fail about 5/6 of the time.
Comments closedSelcuk Disci contrasts a couple of methods for time series forecasting:
It is always hard to find a proper model to forecast time series data. One of the reasons is that models that use time-series data often expose to serial correlation. In this article, we will compare k nearest neighbor (KNN) regression which is a supervised machine learning method, with a more classical and stochastic process, autoregressive integrated moving average (ARIMA).
We will use the monthly prices of refined gold futures(XAUTRY) for one gram in Turkish Lira traded on BIST(Istanbul Stock Exchange) for forecasting. We created the data frame starting from 2013. You can download the relevant excel file from here.
Click through for the demonstration. H/T R-Bloggers.
Comments closedJasmine Nettiksimmons and Molly Davies explain what estimators are:
What makes a good estimator? What is an estimator? Why should I care? There is an entire branch of statistics called Estimation Theory that concerns itself with these questions and we have no intention of doing it justice in a single blog post. However, modern effect estimation has come a long way in recent years and we’re excited to share some of the methods we’ve been using in an upcoming post. This will serve as a gentle introduction to the topic and a foundation for understanding what makes some of these modern estimators so exciting.
Read on for a very nice introduction to the topic.
Comments closed