|
President's Letter
Dear Members,
February Luncheon Recap
I would like to begin by expressing my sincere appreciation to our February Luncheon presenter, Josiah Parry. Josiah, a Senior Product Engineer at ESRI, delivered an insightful presentation titled “Spatial Econometrics: Lags, Randomness, and Regression.”
During his talk, Josiah highlighted the challenges associated with spatial data and emphasized the necessity of specialized analytical methods. He explained how spatial econometrics provides a robust framework for identifying spatial patterns and dependencies, offering valuable insights into spatial data analysis.
In addition to his engaging presentation, Josiah also shared several key references:
Introduction to Spatial Econometrics by James LeSage and Robert Pace (2009)
Modern Spatial Econometrics in Practice: A Guide to GeoDa, GeoDaSpace, and PySAL by Luc Anselin and Sergio J. Rey (2014)
He also discussed his R package, sfdep, which serves as a Tidyverse-compatible interface for the spdep package. According to Josiah’s GitHub page:
“sfdep builds on the strong foundation of the spdep package for spatial dependence. It provides an sf- and Tidyverse-friendly interface while introducing new functionalities not available in spdep. The package extensively utilizes list columns to enable this integration.”
Innovations in Data Science
Continuing with the theme of innovation in data science, I would like to highlight some of my recent work in unsupervised learning. As many of us involved in cluster analysis are aware, high-dimensional data often leads to poor cluster partition quality due to the well-known Curse of Dimensionality. This issue, in turn, affects scoring models by reducing their out-of-sample predictive accuracy.
A widely used technique for dimensionality reduction is Principal Component Analysis (PCA). However, PCA has certain limitations, including its restriction to linear transformations and the inevitable loss of information when discarding components to reduce dimensionality.
An alternative and potentially more effective approach is deep learning autoencoders. Autoencoders not only perform dimensionality reduction without information loss but also capture non-linear relationships within the data. Preliminary research suggests that leveraging autoencoders in cluster analysis may lead to higher-quality partitions and improved model performance.
Upcoming Event: Modern Forecasting Methods Conference
Our Modern Forecasting Methods conference will take place at the Amazon offices in Chicago on March 28th. I look forward to seeing everyone there for an exciting discussion on forecasting techniques.
Best regards,
Joe Retzer
President, Chicago Chapter of the ASA
|