Tree-based methods like decision trees and their powerful random forest extensions are one of the most widely used machine learning algorithms. They are easy to use and provide good forecasting performance off the cuff more or less. Another machine learning community darling is the deep learning method, particularly neural networks. These are ultra flexible algorithms with impressive forecasting performance even (and especially) in highly complex real-life environments.
This post is shares:
Convolutional Neural Networks (CNNs from here on) triumph in the field of image processing because they are designed to effectively handle strong spatial dependencies. Simply put, adjacent pixel-values are close to each other, often changing only gradually from one pixel to the next. In a picture where you wear a blue shirt, all the pixels in that area of the picture are blue. You can think of a strong autocorrelated time series, just for spatial data rather than sequential data. This post explains few important concepts related to CNNs: sparsity of connections, parameter sharing, and hierarchical feature engineering.
While linear correlation (aka Pearson correlation) is by far the most common type of dependence measure there are few arguably better ways to characterize\estimate the degree of dependence between variables. This is a fascinating topic I keep coming back to. There is so much for a typical geek to appreciate: non-linear dependencies, should we consider the noise in the data or rather just focus on the underlying process, should we consider the whole distribution or just few moments.
In this post number 6 on correlation and correlation structure I share another dependency measure called “distance correlation”. It has been around for a while now (2009, see references). I provide just the intuition, since the math has little to do with the way distance correlation is computed, but rather with the theoretical justification for its practical legitimacy.
In the field of unsupervised machine learning, similarity and dissimilarity metrics (and matrices) are part and parcel. These are core components of clustering algorithms or natural language processing summarization techniques, just to name a couple.
While at first glance distance metrics look like child’s play, the fact of the matter is that when you get down to business there are a lot of decisions to make, and who likes that? to make matters worse:
After reading this post you will understand concepts like distance metrics, (dis)similarity metrics, and see why it’s fashionable to use kernels as similarity metrics.
Hyper-parameters are parameters which are not estimated as an integral part of the model. We decide on those parameters but we don’t estimate them within, but rather beforehand. Therefore they are called hyper-parameters, as in “above” sense.
Almost all machine learning algorithms have some hyper-parameters. Data-driven choice of hyper-parameters means typically, that you re-estimate the model and check performance for different hyper-parameters’ configurations. This adds considerable computational burden. One popular approach to set hyper-parameters is based on a grid-search over possible values using the validation set. Faster and simpler ways to intelligently choose hyper-parameters’ values would go a long way in keeping the stretched computational cost at a level you can tolerate.
Enter the paper “Random Search for Hyper-Parameter Optimization” by James Bergstra and Yoshua Bengio, suggesting with a straight face not to use grid-search but instead, look for good values completely at random. This is very counterintuitive, for how can a random guesses within some region compete with systematically covering the same region? What’s the story there?
Below I share the message of that paper, along with what I personally believe is actually going on (and the two are very different).
Every so often I read about the kernel trick. Each time I read about it I need to relearn what it is. Now I am thinking “Eran, don’t you have this fancy blog of yours where you write about statistics you don’t want to forget?” and then: “why indeed I do have a fancy blog where I write about statistics I don’t want to forget”. So in this post I explain the “trick” in kernel trick and why it is useful.
Random forests is one of the most powerful pure-prediction algorithms; immensely popular with modern statisticians. Despite the potent performance, improvements to the basic random forests algorithm are still possible. One such improvement is put forward in a recent paper called Local Linear Forests which I review in this post. To enjoy the read you need to be already familiar with the basic version of random forests.
In volatility modelling, a typical challenge is to keep the covariance matrix estimate valid, meaning (1) symmetric and (2) positive semi definite*. A new paper published in Econometrica (citing from the paper) “introduces a novel parametrization of the correlation matrix. The reparametrization facilitates modeling of correlation and covariance matrices by an unrestricted vector, where positive definiteness is an innate property” (emphasis mine). Econometrica is known to publish ground-breaking research, and you may wonder: what is the big deal in being able to reparametrise the correlation matrix?
Deep learning algorithms are increasingly featuring in popular news outlets, large-scale media events and academic conferences. But what makes them so popular? Why now?
I recently published what I hope is an easy read for all of you modern-statistics geeks lovers; explaining the thrust behind this machine-learning class of models.
You can download the two-pager from Significance, specifically here (subscription required).
More good news for the statistical bootstrap. A new paper in the prestigious Econometrica journal makes two interesting points.
This post is inspired by Leo Breiman’s opinion piece “No Bayesians in foxholes”. The saying “there are no atheists in foxholes” refers to the fact that if you are in the foxhole (being bombarded..), you pray! Leo’s paraphrase indicates that when complex, real problems are present, there are no Bayesian to be found.
Random Forests (RF from here onwards) is a widely used pure-prediction algorithm. This post assumes good familiarity with RF. If you are not familiar with this algorithm, stop here and see the first reference below for an easy tutorial. If you used RF before and you are familiar with it, then you probably encountered those “importance of the variables” plots. We start with a brief explanation of those plots, and the concept of importance scores calculation. Main takeaway from the post: don’t use those importance scores plots, because they are simply misleading. Those importance plots are simply a wrong turn taken by our human tendency to look for reason, whether it’s there or it’s not there.
Very!
A distinctive power of neural networks (neural nets from here on) is their ability to flex themselves in order to capture complex underlying data structure. This post shows that the expressive power of neural networks can be quite swiftly taken to the extreme, in a bad way.
What does it mean? A paper from 1989 (universal approximation theorem, reference below) shows that any reasonable function can be approximated arbitrarily well by fairly a shallow neural net.
Speaking freely, if one wants to abuse the data, to overfit it like there is no tomorrow, then neural nets is the way to go; with neural nets you can perfectly map your fitted values to any data shape. Let’s code an example and explain the meaning of this.
This is the fifth post which is concerned with quantifying the dependence between variables. When talking correlations one usually thinks about linear correlation, aka Pearson’s correlation. One serious limitation of linear correlation is that it’s, well.. linear. By construction it’s not useful for detecting non-monotonic relation between variables. Here I share some recent academic research, a new way to detect associations that are not monotonic.
Principal component analysis (PCA from here on) is performed via linear algebra functions called eigen decomposition or singular value decomposition. Since you are actually reading this, you may well have used PCA in the past, at school or where you work. There is a strong link between PCA and the usual least squares regression (previous posts here and here). More recently I explained what does variance explained by the first principal component actually means.
This post offers a matrix approximation perspective. As a by-product, we also show how to compare two matrices, to see how different they are from each other. Matrix approximation is a bit math-hairy, but we keep it simple here I promise. For this fascinating field itself I suspect a rise in importance. We are constantly stretching what we can do computationally, and by using approximations rather than the actual data, we can ease that burden. The price for using approximation is decrease in accuracy (à la “garbage in garbage out”), but with good approximation the tradeoff between the accuracy and computational time is favorable.
