Let’s say you have data containing a categorical variable with 50 levels. When you divide the data into train and test sets, chances are you don’t have all 50 levels featuring in your training set.

This often happens when you divide the data set into train and test sets according to the distribution of the outcome variable. In doing so, chances are that our explanatory categorical variable might not be distributed exactly the same way in train and test sets – so much so that certain levels of this categorical variable are missing from the training set. The more levels there are to a categorical variable, it gets difficult for that variable to be similarly represented upon splitting the data.

Take for instance this example data set (train.csv + test.csv) which contains a categorical variable var_b that takes 349 unique levels. Our train data has 334 of these levels – on which the model is built – and hence 15 levels are excluded from our trained model. If you try making predictions on the test set with this model in R, it throws an error: factor var_b has new levels 16060, 17300, 17980, 19060, 21420, 21820,
25220, 29340, 30300, 33260, 34100, 38340, 39660, 44300, 45460

If you’ve used R to model generalized linear class of models such as linear, logit or probit models, then chances are you’ve come across this problem – especially when you’re validating your trained model on test data.

The workaround to this problem is in the form of a function, remove_missing_levels that I found here written by pat-s. You need magrittr library installed and it can only work on lm, glm and glmmPQL objects.

Once you’ve sourced the above function in R, you can seamlessly proceed with using your trained model to make predictions on the test set. The code below demonstrates this for the data set shared above. You can find these codes in one of my github repos and try it out yourself.

I’ve been stuck for about a week at the 52nd percentile among 3400+ Kagglers taking part in the competition. I’ve been told that Kaggle Kernels and discussion boards are helpful when you’re stuck or if you need to learn some practical data science that can’t be gleaned from books or tutorials.

One such discussion thread looks like this:

This person going by the pseudonym Schoolpal is currently killing it on the leaderboard and I’m eagerly looking forward to this person’s code once the competition ends in less than 24 hours. If you’re interested too, follow this discussion here.

Cheers!

Update:

This Schoolpal, as mentioned earlier, finally came in second and shared their approach here.

The most conventional approach to determine structural breaks in longitudinal data seems to be the Chow Test.

From Wikipedia,

The Chow test, proposed by econometrician Gregory Chow in 1960, is a test of whether the coefficients in two linear regressions on different data sets are equal. In econometrics, it is most commonly used in time series analysis to test for the presence of a structural break at a period which can be assumed to be known a priori (for instance, a major historical event such as a war). In program evaluation, the Chow test is often used to determine whether the independent variables have different impacts on different subgroups of the population.

As shown in the figure below, regressions on the 2 sub-intervals seem to have greater explanatory power than a single regression over the data.

For the data above, determining the sub-intervals is an easy task. However, things may not look that simple in reality. Conducting a Chow test for structural breaks leaves the data scientist at the mercy of his subjective gaze in choosing a null hypothesis for a break point in the data.

Instead of choosing the breakpoints in an exogenous manner, what if the data itself could learn where these breakpoints lie? Such an endogenous technique is what Bai and Perron came up with in a seminal paper published in 1998 that could detect multiple structural breaks in longitudinal data. A later paper in 2003 dealt with the testing for breaks empirically, using a dynamic programming algorithm based on the Bellmanprinciple.

I will discuss a quick implementation of this technique in R.

Brief Outline:

Assuming you have a ts object (I don’t know whether this works with zoo, but it should) in R, called ts. Then implement the following:

An illustration

I started with data on India’s rice crop productivity between 1950 (around Independence from British Colonial rule) and 2008. Here’s how it looks:

You can download the excel and CSV files here and here respectively.

Here’s the way to go using R:

Voila, this is what you get:

The dotted vertical lines indicated the break dates; the horizontal red lines indicate their confidence intervals.

This is a quick and dirty implementation. For a more detailed take, check out the documentation on the R package called strucchange.

This was in the pipeline for quite some time now. I have been waiting for his lectures on a platform such as EdX or Coursera, and the day has arrived. You can enroll and start with week 1’s lectures as they’re live now.

This course is taught by none other than Dr. Yaser S. Abu – Mostafa, whose textbook on machine learning, Learning from Data is #1 bestseller textbook (Amazon) in all categories of Computer Science. His online course has been offered earlier over here.

Teaching

Dr. Abu-Mostafa received the Clauser Prize for the most original doctoral thesis at Caltech. He received the ASCIT Teaching Awards in 1986, 1989 and 1991, the GSC Teaching Awards in 1995 and 2002, and the Richard P. Feynman prize for excellence in teaching in 1996.

Live ‘One-take’ Recordings

The lectures have been recorded from a live broadcast (including Q&A, which will let you gauge the level of CalTech students taking this course). In fact, it almost seems as though Abu Mostafa takes a direct jab at Andrew Ng’s popular Coursera MOOC by stating the obvious on his course page.

A real Caltech course, not a watered-down version

Again, while enrolling note that this is what Abu Mostafa had to say about the online course: “A Caltech course does not cater to short attention spans, and it may not provide instant gratification…[like] many MOOCs out there that are quite simple and have a ‘video game’ feel to them.” Unsurprisingly, many online students have dropped out in the past, but some of those students who “complained early on but decided to stick with the course had very flattering words to say at the end”.

Prerequisites

Basic probability

Basic matrices

Basic calculus

Some programming language/platform (I choose Python!)

If you’re looking for a challenging machine learning course, this is probably one you must take.

I love probability and have solved countless problems on probability ever since I learned math

…and yet I’ve never coded up probabilistic models!

The assignments and project work for this course are to be implemented in Python!

You don’t need to have prior experience in either probability or inference, but you should be comfortable with basic Python programming and calculus.

WHAT YOU’LL LEARN
– Basic discrete probability theory
– Graphical models as a data structure for representing probability distributions
– Algorithms for prediction and inference
– How to model real-world problems in terms of probabilistic inference

The course started on September 12, is 12-weeks long and is structured in the following manner:

Week 1 (9/12 – 9/16): Introduction to probability and computation
A first look at basic discrete probability, how to interpret it, what probability spaces and random variables are, and how to code these up and do basic simulations and visualizations.

Week 2 (9/19 – 9/23):Incorporating observations
Incorporating observations using jointly distributed random variables and using events. Three classic probability puzzles are presented to help elucidate how to interpret probability: Simpson’s paradox, Monty Hall, boy or girl paradox.

Week 3 (9/26 – 9/30):Introduction to inference, structure in distributions, and information measures
The product rule and inference with Bayes’ theorem. Independence: A structure in distributions. Measures of randomness: entropy and information divergence. Mutual information.

Week 4 (10/3 – 10/7):Expectations, and driving to infinity in modeling uncertainty
Expected values of random variables. Classic puzzle: the two envelope problem. Probability spaces and random variables that take on a countably infinite number of values and inference with these random variables.

Week 5 (10/10 – 10/14):Efficient representations of probability distributions on a computer
Introduction to undirected graphical models as a data structure for representing probability distributions and the benefits/drawbacks of these graphical models. Incorporating observations with graphical models.

Week 6 (10/17 – 10/21):Inference with graphical models, part I
Computing marginal distributions with graphical models in undirected graphical models including hidden Markov models..

Week 7 (10/24 – 10/28):Inference with graphical models, part II
Computing most probable configurations with graphical models including hidden Markov models.

Week 8 (10/31 – 11/4):Introduction to learning probability distributions
Learning an underlying unknown probability distribution from observations using maximum likelihood. Three examples: estimating the bias of a coin, the German tank problem, and email spam detection.

Week 9 (11/7 – 11/11):Parameter estimation in graphical models
Given the graph structure of an undirected graphical model, we examine how to estimate all the tables associated with the graphical model.

Week 10 (11/14 – 11/18):Model selection with information theory
Learning both the graph structure and the tables of an undirected graphical model with the help of information theory. Mutual information of random variables.

Week 11 (11/21 – 11/25):Final project
Final project assigned

Week 12 (11/28 – 12/2):Final project

I’m SO taking this course. Hope this interests you as well!

This was a hackathon + workshop conducted by Analytics Vidhya in which I took part and made it to the #1 on the leaderboard. The data set was straight-forward and quite clean with only a minor need for missing value treatment. This post will might be useful for people who want a walk-through on the steps involving data munging and developing machine-learned models.

The workshop ended with a basic hackathon with data given on age, education, working class, occupation, marital status and gender of individuals and one had to predict the income bracket of these individuals.

I’ve posted the data and my code and solutions in this GitHub repo. An IPython Notebook has also been shared.

I approached the problem first by attempting some feature engineering (other than missing value treatment) on the data, and then ran a basic logistic classifier and a random forest classifier. However it turned out that these models performed better without feature engineering, which shows the dataset was already quite clean and informative to begin with for this competition.

I later attempted gradient boosting with parameter tuning to maximizing scores.

Here’s a quick example case for implementing one of the simplest of learning algorithms in any machine learning toolbox – Linear Regression. You can download the IPython / Jupyter notebook here so as to play around with the code and try things out yourself.

I’m doing a series of posts on scikit-learn. Its documentation is vast, so unless you’re willing to search for a needle in a haystack, you’re better off NOT jumping into the documentation right away. Instead, knowing chunks of code that do the job might help.