data

Sometimes I want to extract text from HTML for processing, but I don't want to lose the context. This is useful in NLP with HTML because sometimes the context (is it emphasised, or in a header or list item) may be relevant for capturing information. The common approaches of converting HTML to text are lossy; you strip away the HTML and lose the context. Instead we're going to do create a source mapping using the Python HTML Parser, for each character of the output text we can find the corresponding character of the HTML that generated it.

We’re going to build up a simple Deep Learning API inspired by fastai on Fashion MNIST from scratch. Humans can only fit so many things in their head at once (somewhere between 3 and 7); trying to grasp all the details of the training loop at once is difficult, especially as we add more features to it. The right abstractions can make this much easier by only having to think about what we’re changing in the interface.

Chapter 4 of the fastai book covers how to build a Neural Network for distinguishing 3s and 7s on MNIST from scratch. We’re going to do a similar thing but instead of building the neural network from the ground up we’re going to use fastai’s layered API to build it top down. We’ll start with the high level API to train a dense neural network in a few lines. Then we’ll redo the problem going deeper and deeper into the API.

Going from batch processing to near-real time applications is a big conceptual leap for data scientists. Data scientists are often familiar with big SQL analytics databases and can run a batch process weekly or daily. However to get good performance in some applications requires aggregating information that changes more quickly than batch processes can handle. There is normally some lag time between an event being processed and being ingested into an analytics database (and this lag time can vary across the data) which limits the batch approach.

I've been trying to use pagination with the Internet Archive's CDX for the Wayback Machine but have been getting lots of empty results. The reason is that filters are applied after pagination, and so when using tight filters almost all records will be empty. Using the CDX API Following the documentation it's easy enough to make a query (and similar, but not the same, as Common Crawl's CDX Index). Here's some logic to get captures from en.

I'm currently developing a workflow for extracting data from captures of web pages, leveraging large archives like Common Crawl and the Wayback Machine. A pain point in extracting data from external sources is you have to reverse engineer the schema, and quite often it breaks on certain data. To reduce the pain of hitting these errors it makes sense to cache the results to reduce the time of re-extracting the data; especially if it's doing something expensive like parsing a large HTML document.

When calculating the averages of lots of different groups it doesn't make sense to treat the groups as independent, but to pool information across groups, especially on groups with little data. One way to do this is to build a model on covariates of the groups or on categorical embeddings to use information from other observations to inform that observation. Surprisingly Stein's Paradox says if we're trying to minimise the root mean square error, have only one observation per group, and they're all normally distributed with the same standard deviation we're always better off shrinking the means towards a common point than treating them separately (even if they're completely unrelated quantities!

Stein's Paradox states that when trying to estimate the 3 or more means of normally distributed data together, it's always better (on average) to shrink the estimates. Specifically if you've got p independent normally distributed variables \(X_i \sim N(\theta_i, 1) ;\, i=1,\ldots,p\) the best estimates for minimising the mean squared error of all the estimates isn't the values themselves \(X\), and the James-Stein estimator is better (has strictly lower risk).

Estimating a Bernoulli Probability Suppose we want to know the probability of an event occurring; it could be a customer converting, a person contracting a disease or a student passing a test. This can be represented by a Bernoulli Distribution, where each draw is an independent random variable \( \gamma_i \sim {\rm Bernoulli}(\theta) \). The only possible values are failure (represented by 0) with probability \(\theta\) and success (represented by 1) with probability \(1-\theta\) (although the labels are completely arbitrary and we can switch them by setting \(\eta_i = 1 - \gamma_i\) then \(\eta_i \sim {\rm Bernoulli}(1-\theta) \)).

High cardinality categorical data are tricky for machine learning models to deal with. A linear model tries to estimate a different coefficient for every category, treating them as totally independent. There are tools like hierarchical models that can encode some cross-correlations but (to my knowledge) they don't scale well to large datasets. A tree model will try to estimate different coefficients on groups of categories (based on the order they are sorted in), but for many categories there is no canonical order (like locations on a map).

This is the third on a series of articles showing the basics of building models in Stan and accessing them in R. Now that we can specify a linear model and fit it in with formula syntax, and specify priors for the model, it would be useful to be able to make predictions with it. In principle making predictions from our linear model \( y \sim N(\alpha + \beta x, \sigma)\) is easy; to make point predictions we take central estimates of the coefficients \(\hat{\alpha}\) and \(\hat{\beta}\) and estimate \( y \approx \hat{\alpha} + \hat{\beta} x\).

This is the second on a series of articles showing the basics of building models in Stan and accessing them in R. In the previous article I showed how to specify a simple linear model with flat priors in Stan, and fit it in R with a formula syntax. In this article we extend this to specify priors; defaulting to general weakly informative priors but allowing use of specific priors.

I wanted to fit a Bayesian Tobit model, but I couldn't find one (probably because I didn't know how to look). So I decided to build one in Stan, which I had never used before. This article is the first in a series showing how I got there; this one builds a linear model in Stan and makes it useable from R using formula syntax, then next we add priors to the model, make model predictions from R, and then handle censored values with Tobit regression.

In Bayesian statistics you have to choose a prior distribution for the parameters to combine with the data to get a posterior distribution. Choosing a tight prior, assuming that the parameters should live in a particular space, reduces the impact of the data on the posterior estimates. This is just like regularisation in machine learning where adding a penalty to the loss function prevents over-fitting. This is more than just an analogy, and this article will explore a couple of cases with constant regression and classification.

When understanding how a binary prediction depends on a continuous input I find a very useful way is to bin it into quantiles and plot the average probability. For example here's a plot using the iris dataset showing how the probability that a flower is a "virginica" and not a "versicolor" changes with the sepal width of the flower. This kind of plot can show nonlinearities and indicate how we should include this variable in a logistic regression.

Interpretable models are incredibly useful for getting users to trust your model (and understand when not to trust it). In projects helping subject matter experts make decisions it helps them see what's going into it, and what adjustments need to be made. However a large part of what makes a model interpretable is not just its form, but how it's presented. Reparameterising it in terms that make sense from a decision makers point of view can make all the difference in gaining insight, trust and adoption.

Many analytics codebases consist of a pipeline of steps, doing things like getting data, extracting features, training models and evaluating results and diagnostics. The best way to structure the code isn't obvious and if you're having trouble importing files, getting module not found errors or are tinkering with PYTHONPATH it's likely you haven't got it right. A way I've seen many data analytics processing pipelines structured is with a series of numbered steps:

Gradient descent is an effective algorithm for finding the local extrema of functions, and the global extrema of convex functions. It's very useful in machine learning for fitting a model from a family of models by finding the parameters that minimise a loss function. However sometimes you have extra constraints; I recently worked on a problem where the maximum value of the fitted function had to occur at a certain point.

Modern Javascript web frameworks often embed the data used to render each webpage in the HTML. This means an easy way of extracting data is capturing the string representation of the object with a pushdown automoton and then parsing it. Sometimes Python's json.loads won't cut it for dynamic JSON; one option is demjson but another much faster option is chompjs. Chompjs converts a javascript string into something that json.loads. It's a little less strict than demjson; for example {"key": undefined} will be converted by chompjs.

The internet is full of useful information just waiting to be collected, processes and analysed. While getting started scraping data from the web is straightforward, it's easy to tangle the whole process together in a way that makes it fragile to failure, or hard to change with requirements. And the internet is inconsistent and changing; in anything but the smallest scraping projects you're going to run into failures. I find it useful to conceptually break web scraping down into four steps; selecting the data to retrieve, fetching the data from the internet, extracting structured data from the raw responses, and watching the process to make sure it's functioning correctly.

Suppose you've got two dice; what's the probability the sum of their rolls will add up to 4? You simply look at all the ways the values of the dice could sum to 4 (e.g. 1 and 3, 2 and 2 or 1 and 4), and add up their probabilities (in this case each is 1/36, so totalling 3/36 or 1/12). You could use a probability square to visualise this calculation.

What sample size do you need to capture an effect with 95% confidence 80% of the time? For a normal/binomial distribution the answer is roughly \( \left(2.8 \frac{\sigma}{\epsilon}\right)^2 \), where \( \sigma \) is the standard deviation of the data and \( \epsilon \) is the size of the effect. The ratio \( \frac{\sigma}{\epsilon} \) says the smaller the effect size is relative to the variability in the data the larger the sample size you will need.

The demand curve in economics represents the relationship between price and quantity sold. It's generally not possible to know the demand curve without varying prices to measure it. But if you lower a price and then raise it again you don't always get back to your original volume - you get price hysteresis. Hysteresis is where the value of a quantity depends on how you got there. Pricing hysteresis is about that the quantity of goods sold isn't dependent just on the price today, but on previous prices too.

When running an A/B test the sample sizes can seem insane. For example to observe a 2 percentage point uplift on a 60% conversion rate requires over 9,000 people in each group to get the standard 95% confidence level with 80% power. If you've only got less than 18,000 customers you can reach, which is very common in businesss to business settings, it's impossible to conduct this test. But if you look in terms of the outcomes, doing an A/B test on a few hundred users may actually greatly increase your outcomes.

Suppose you have a long standing heart condition, and are considering undergoing a surgical procedure that could alleviate the procedure, but has its own set of risks. You happen to have a good friend who's a statistician at the hospital you're being seen at and can get you historical frequencies of complications. However she asks you what reference set do you want to use? Do you want complications related to just the specific procedure treating your condition, or for all procedures on that area of the heart for a variety of conditions?

The normal distribution is used throughout statistics, because of the Central Limit Theorem it occurs in many applications, but also because it's computationally convenient. The expectation value of the normal distribution is the mean, which has many nice arithmetic properties, but the drawback of being sensitive to outliers. When discussing constant models I noted that the minimiser of the Lᵖ error is a generalisation of the mean; for \( p = 2 \) it's the mean, for \( p = 1 \) it's the median, and for \( p = \infty \) it's the midrange (half way betwen the maximum and minimum points).

When evaluating binary classifier (e.g. will this user convert?) the most obvious metric is accuracy; what's the probability a random prediction is correct. One issue with this metric is if 90% of the cases are one class a high accuracy isn't really impressive; you need to contrast it with a constant model predicting the most frequent class. More subtly it's not a very sensitive measure, by measuring cross-entropy of predicted probabilities you get a much better idea of how well your model is working.

As a professional your reputation is very important to your career success. To get people to offer you work, to pay for your advice or to buy products from you they need to trust that you will deliver them value. The most common heuristic for this is your reputation; what other people say about you, what you have done and what certifications you have. It's important that your reputation is very specific to the kind of work you want others to buy.

In mathematics it's surprising how often something that's obvious (or trivial) to someone else can be revolutionary (or weeks of work) to someone else. I was looking at the annoy (Approximate Nearest Neighbours, Oh Yeah) library and saw this comment: Cosine distance is equivalent to Euclidean distance of normalized vectors I hadn't realised it at all, but once the claim was made I could immediately verify it. Given two vectors u and v their distance is given by the length of the vector between them: \( d = \| u - v \| = \sqrt{(u - v) \cdot (u - v)} \).

Can we have an interpretable model that has as good performance as blackbox models like gradient boosted trees and neural networks? In a 2020 Empirical Methods for Natural Language Processing Keynote, Rich Caruana says yes. He calls interpretable models glassbox machine learning, in contrast to blackbox machine learning. It is models in which a person can explicitly see how they work, and follow the steps from inputs to outputs. This interpretability is subtly different from explainable (explainable to who?

When people access products online their behaviour gives lots of information about both the people and the products. This information deeply enriches understanding of how to better serve your customers, how your products are related to each other and can help answer deeper questions about them. However you need to find a way to unlock the information. Using behavioural information can greatly improve modelling on the tabular data in your database.

Analytics projects are messy. It's rarely clear at the start how to frame the business problem, whether a given approach will actually work, and if you can get it adopted by your partners. However once you have a framing the modelling part can be iterated on quickly by structuring the project like a Kaggle Competition. The modelling part of analytics projects will go smoothly only if you have clear evaluation criteria.

When looking at Pandas dataframes in a Jupyter notebook it can be hard to find what you're looking for in a big mess of numbers. Something that can help is formatting the numbers, making them shorter and using graphics to highlight points of interest. Using Pandas style you can make the story of your dataframe standout in a Jupyter notebook, and even export the styling to Excel. The Pandas style documentation gives pretty clear examples of how to use it.

I started regularly writing this website to get better at writing, to build a portfolio and share my learnings. Because of this I haven't been focused on building an audience or looking at analytics. However now I've been writing continuously for 6 months I'd see if I learned anything interesting from looking at Google Analytics. I installed Google Analytics a couple of weeks ago on the website to see how people are actually viewing my site.

A couple years ago I built whatcar.xyz which predicts the make and model of Australian cars. It was built mainly with externally sourced data and so only works sometimes, under good conditions. To make it better I've started collecting my own training data. External data sources are extremely convenient for training a model as they can often be obtained much more cheaply than curating your own data. But the data will almost always be different to what you are actually performing inference on, and so you're relying on a certain amount of generalisation.

Experiments reveal the relationship between inputs and outcomes. With statistical methods you can often, with enough observations, tell whether there's a strong relationship or if it's just noise. However it's much harder to know how generally the relationship holds, but it's essential for making decisions. Suppose you're testing two alternate designs for a website. One has a red and green button with a santa hat and bauble, and the other has a blue button.

I've spent the last few hours looking at the Gridded Population of the World which consistenly estimates the population density consistent with national censuses and population registers. This would have been a massive job to compile and is really interesting to look at. You can immediately see a strip through the north of India, Pakistan and Bangladesh that is incredibly dense. The north-east of China and the island of Java in Indonesia are also very dense.

When making changes to a new model training pipeline I find it really useful to understand the dataflow. Analytics workflows are done as a series of transformations, taking some inputs and producing some outputs (or in the case of mutation; an input is also an output). Seeing this dataflow helps give a big picture overview of what is happening and makes it easier to understand the impact of changes. Generally you can view the process as a directed and (hopefully) acyclic graph.

The state government of Victoria, Australia has recently announced a plan on how to respond to the current Covid-19 pandemic. Based on epidemiological modelling they have set to reduce restrictions based on 14 day averages of new case numbers. If the 14 day average daily new cases are 30-50 in 3 weeks they will reduce restrictions; if they are below 5 a month after that they will reduce restrictions again.

Understanding the spread of infectious disease is very important for policies around public health. Whether it's the seasonal flu, HIV or a novel pandemic the health implications of infectious diseases can be huge. A change in decision can mean saving thousands of lives and relieving massive suffering and related economic productivity losses. The SIR model is a model that is simple, but captures the underlying dynamics of how quickly infectious diseases spread.

The starting point for an analysis is often summary statistics, such as the mean or the median. For some of these you're going to want it more precisely, more timely or cut by thinner segments. When the data gets too volatile to report on it's a good time to re-frame the descriptive statistics as a predictive problem. Businesses often have a lot of reporting around important metrics cut by key segments.

A while ago I came across Cynthia Rudin through their work on the FICO Explainable Machine Learning Challenge. Her team got an honourable mention and she wrote an opinion with Joanna Radin on explainable models. I think the article was hyperbolic on claiming interpretable models always work as well as black box models. On the other hand I only came across her because of this article, so taking an extreme viewpoint in the media is a good way to get attention.

Sometimes you want to classify documents, but you don't have an existing classification. Building a classification that is mutually exclusive and completely exhaustive is actually very hard. Topic modelling is a great way to quickly get started with a basic classification. Creating a classification may sound easy until you try to do it. Think about novels; is a Sherlock Holmes novel a mystery novel or a crime novel (or both)? Or do we go more granular and call it a detective novel, or even more specifically a whodunit?

A coarse geocoder takes a human description of a large area like a city, area or country and returns the details of that location. I've been looking into the source of the excellent Placeholder (a component of the Pelias geocoder) to understand how this works. The overall approach is straightforward, but it takes a lot of work to get it to be reliable. A key component geocoder is a gazetteer that contains the names of locations.

Placeholder is a great library for Coarse Geocoding, and I'm using it for finding locations in Australia. In my application I want to get the location to a similar level of granularity; however the input may be for a higher level of granularity. Placeholder doesn't directly provide a method to do this, but you can use their SQLite database to do it. For example to find the largest locality for East Gippsland, with Who's On First id 102049039, you can use the SQL.

Sometimes you may want to experiment with sessions and need to hand-roll your own in SQL. There's a good mode blog on how to do this. If you're using Postgres or Greenplum you may be able to use Apache Madlib's Sessionize for the basic case. This blog post will give a very brief summary of how to do this with some examples in Presto/Athena. The idea of a session is to capture a continuous unit of user activity.

Differentiation is the process of creating a local linear approximation of a function. This is useful because linear functions are very well understood and efficient to work with. One application of them is gradient descent, often used for fitting models in machine learning. In this context a function is something that maps between coordinate spaces. For example consider an image classifier that takes a 128x128 pixel image with three channels for colours (Red, Green, Blue) and returns a probability that the image contains a cat and the probability that the image contains a dog.

A challenge of data analytics is that the data can change as well as the code. The systems producing and collecting data are often changed and can lead to missing or corrupt data. These can easily corrupt reports and machine learning systems. Worst of all the data may be lost permanently. So if you're going to use some data it's important to check the data regularly to catch the worst kind of mistakes as early as possible.

You don't need a lot of data to prove a point. People often think statistics requires big expensive datasets that cost a lot to acquire. However in relatively unexplored spaces a small amount of data can have high yield in changing a decision. I've been working on some problems around web sessionisation. The underlying model is that when someone visits your website they may come at different times for different reasons.

Tracking a metric can help to drive dramatic improvements. When your team is focused on a metric you can test what has impact and quickly optimise it. However for this to work it's important to be something you can actually impact. When people start looking for a metric to track they want to look for things that have a direct impact on the business, such as revenue, share price or customer satisfaction.

When training machine learning models typically you get a training dataset for fitting the model and a test dataset for evaluating the model (on small datasets techniques like cross-validation are common). You typically assume the performance on your chosen metric on the test dataset is the best way of judging the model. However it's really easy for systematic biases or leakage to creep into the datasets, meaning that your evaluation will differ significantly to real world usage.

Deep Neural Networks have transformed dealing with unstructured data like images and text, making totally new things possible. However they are difficult to train, require a large amount of relevant training data, are hard to interpret, hard to debug and hard to refine. I think for these reasons there's a lot of space to use neural networks as a building block for extracting structured data for less parameterised models. Josh Tenenbaum gave an excellent keynote at ACL 2020 titled Cognitive and computational building blocks for more human-like language in machines.

The traditional way to train an NER model on a new domain is to annotate a whole bunch of data. Techniques like active learning can speed this up, but especially neural models with random weights require a ton of data. A more modern approach is to take a large pretrained NER model and fine tune it on your dataset. This is the approach of AdaptaBERT (paper), using BERT. However this takes a large amount of GPU compute and finicky regularisation techniques to get right.

Many documents available on the web have meaningful markup. Headers, paragraph breaks, links, emphasis and lists all change the meaning of the text. A common way to deal with HTML documents in NLP is to strip away all the markup, e.g. with Beautiful Soup's .get_text. This is fine for a bag of words approach, but for more structured text extraction or language model this seems like throwing away a lot of information.

Modern Javascript web frameworks often embed the data used to render each webpage in the HTML. This means an easy way of extracting data is capturing the string representation of the object with a pushdown automoton and then parsing it. Python's inbuilt json.loads is effective, but won't handle very dynamic Javascript, but demjson will (another, much faster alternative is Chompjs. The problem shows up when using json.loads as the following obscure error:

Downloading files can often be a bottleneck in a data pipeline because network I/O is slow. A really simple way to handle this is to run multiple downloads in parallel accross threads. While it's possible to deal with the unused CPU cycles using asynchronous processing, in Python it's generally easier to throw more threads at it. Using multiprocessing can be very simple if you can turn make the processing occur in a pure function or object method, and both the variables are results are picklable.

Data pipelines can often be thought of as a chain of pure functions passing data between them, even if they are not implemented that way. As long as you can access the sources you can always rerun the pipeline from end to end. However if the processing takes a long time to run it can be convenient to cache some step as a sink, especially if steps are likely to fail.

I am trying to extract Australian Job Postings from Web Data Commons which extracts structured data from Common Crawl. I previously came up with a SPARQL query to extract the Australian jobs from the domain, country and currency. Unfortunately it's quite slow, but we can speed it up dramatically by replacing it with a similar script in grep. With a short grep script we can get twenty thousand Australian Job Postings with metadata from 16 million lines of compressed nquad in 30 seconds on my laptop.

Sometimes you have some description of a location and want to work out where it is. This is called geocoding; if you just want to know what state or country it's in it's called coarse geocoding. I found that while many structured JobPostings contain a country some have it as a description rather than a country code, and some put the location in other fields. We can often find the country using geocoding.

I am trying to understand how the JobPosting schema is used in Web Data Commons structured data extracts from Common Crawl. I wrote a lot of ad hoc Python to get usage statistics on JobPosting. However SPARQL is a tool that makes it much easier to answer these kinds of questions. After reading in the graphs individually they can be combined into a rdflib.Dataset so we can query them all together.

The Web Data Commons extracts structured RDF Data from about one monthly Common Crawl per year. These contain a vast amount of structured information about local businesses, hostels, job postings, products and many other things from the internet. Python's RDFLib can read the n-quad format the data is stored in, but by default requires reading all of the millions to billions of relations into memory. However it's possible to process this data in a streaming fashion allowing it to be processed much faster.

Common Crawl builds an open dataset containing over 100 billion unique items downloaded from the internet. There are petabytes of data archived so directly searching through them is very expensive and slow. To search for pages that have been archived within a domain (for example all pages from wikipedia.com) you can search the Capture Index. But this doesn't help if you want to search for paths archived across domains. For example you might want to find how many domains been archived, or the distribution of languages of archived pages, or find pages offered in multiple languages to build a corpus of parallel texts for a machine translation model.

Common Crawl builds an open dataset containing over 100 billion unique items downloaded from the internet. You can search the index to find where pages from a particular website are archived, but you still need a way to access the data. Common Crawl provides the data in 3 formats: If you just need the text of the internet use the WET files If you just need the response metadata, HTML head information or links in the webpage use the WAT files If you need the whole HTML (with all the metadata) then use the full WARC files The index only contains locations for the WARC files, the WET and WAT files are just summarisations of it.

Common Crawl builds an open dataset containing over 100 billion unique items downloaded from the internet. Every month they use Apache Nutch to follow links accross the web and download over a billion unique items to Amazon S3, and have data back to 2008. This is like what Google and Bing do to build their search engines, the difference being that Common Crawl provides their data to the world for free.

Different industries have different ways of distinguishing seniority in a job title. Is a HR Officer more senior than a HR Administrator? Is a PHP web developer more skilled than a PHP developer? How different is a medical sales executive to general sales roles? Using the jobs from Adzuna Job Salary Predictions Kaggle Competition I've found common job titles and can use the advertised salary to help understand them. Note that since the data is from the UK from several years ago a lot of the details aren't really applicable, but the techniques are.

I've been trying to find near duplicate job ads in the Adzuna Job Salary Predictions Kaggle Competition. Job ads can be duplicated because a hirer posts the same ad multiple times to a job board, or to multiple job boards. Finding exact duplicates is easy by sorting the job ads or a hash of them. But the job board may mangle the text in some way, or add its own footer, or the hirer might change a word or two in different posts.

We've found pairs of near duplicate texts in 400,000 job ads from the Adzuna Job Salary Predictions Kaggle Competition. When we tried to extracted groups of similar ads by finding connected components in the graph of similar ads. Unfortunately with a low threshold of similarity we ended up with a chain of ads that were each similar, but the first and last ad were totally unrelated. One way to work around this is to find cliques, or a group of job ad were every job ad is similar to all of the others.

I needed to design some heuristic thresholds for grouping together items. In my first step attempt I iteratively tried to guess the thresholds by trying them on different examples. This was directionally useful but as I refined the thresholds I had to keep going back to check whether I had broken earlier examples. To improve this I used a spreadsheet as a rough annotation tool. There are various tools for data entry like org mode tables in Emacs, or you can use a spreadsheet interface in R with data.

When you have behavioural data between actors and events you naturally get a bipartite graph. For example you can have the actors as customers and events as products that are purchased, or the actors as users of a website and the events as videos that are viewed, or the actors as members of a forum and the events as posts they comment on. One of the ways to represent this is to relate actors by the number of events they both participate in.

Suppose you're running a website with tens of thousands of different products, and no satisfactory way to group them up. Even a mediocre clustering can really help bootstrap your understanding. You can use the clusters to see new patterns in the data, and you can manually refine the clusters much more easily than you can make them. There are many techniques to cluster structured data or even detect them as communities in the graph of interactions with your users.

I've found pairs of near duplicates texts in the Adzuna Job Salary Predictions Kaggle Competition using Minhash. One thing that would be useful to know is what the common sections of the ads are. Typically if they have a high 3-Jaccard similarity it's because they have some text in common. The most asymptotically efficient to find the longest common substring would be to build a suffix tree, but for experimentation the heuristics in Python's DiffLib work well enough.

My first instinct when dealing with a new problem is to try to find a complex technique to solve it. However I've almost always found it more useful to start with a simple model before trying something more complex. You gain a lot from trying simple models and the cost is low. Even if they're not enough to solve the problem (which they can be) they will often give a lot of information about the problem which will set you up for later techniques.

Suppose that you flip a fair coin 10 times, how many heads will you get? You'd think it was close to 5, but it might be a bit higher or lower. If you only got 7 heads would you reconsider you assumption the coin is fair? What if you got 70 heads out of 100 flips? This might seem a bit abstract, but the inverse problem is often very important. Given that 7 out of 10 people convert on a new call to action, can we say it's more successful than the existing one that converts at 50%?

Dealing with thousands of different items is difficult. When you've got a couple of dozen you can view them together, but as you get into the hundreds, thousands and beyond it becomes necessary to group items to make sense of them. For example if you've got a list of customers you might group them by state, or by annual spend. But sometimes it would be useful to split them into a few groups using some heuristic criteria; clustering is a powerful technique to do this.

A decision tree is a series of conditional rules leading to an outcome. When stated as a chain of if-then-else rules it can be really hard to understand what is going on. If the number of dimensions and cutpoints is relatively small it can be useful to visualise on a grid to understand the tree. Decision trees are often represented as a hierarchy of splits. Here's an example of a classification tree on Titanic survivors.

When you've got a timeseries that doesn't have a timezone attched to it the natural question is "what timezone is this data from?" Sometimes it's UTC, sometimes it's the timezone of the server, otherwise it could be the timezone of one of the locations it's about (and it may or may not change with daylight savings). When it's people's web activity there's a simple heuristic to check this: the activity will be minimum between 3am and 5am.

A very useful open dataset the Australian Government provides is the Geocoded National Address File (G-NAF). This is a database mapping addresses to locations. This is really useful for applications that want to provide information or services based on someone's location. For instance you could build a custom store finder, get aggregate details of your customers, or locate business entities with an address, for example ATMs. There's another open and editable dataset of geographic entities, Open Street Map (and it has a pretty good open source Android app OsmAnd).

Sometimes I get out pipe tables in Emacs that I want to convert into a CSVto put somewhere else. This is really easy with regular expressions. I often get data output from an SQL query like this text | num | value --------------+------+------------- Some text | 0.3 | 0.2 Rah rah | 7 | 0.00123(2 rows) Running sed 's/\(^ *\| *|\|(.*\) */,/g' gives: ,text,num,value --------------+------+------------- ,Some text,0.3,0.2 ,Rah rah,7,0.00123, I can delete the divider and then use as a CSV.

Generally when combining datasets you want to join them on some key. But sometimes you really want a range lookup like Excel's VLOOKUP. A common example is binning values; you want to group values into custom ranges. While you could do this with a giant CASE statement, it's much more flexible to specify in a separate table (for regular intervals you can do it with some integer division gymnastics). It is possible to implement VLOOKUP in SQL by using window functions to select the right rows.

Behavioural data can illuminate the structure of the underlying actors. For example looking at which products customers buy can help understand how both the products and customers interact. The same idea can apply to people who attend events, watch the same movie, or have authored a scientific paper together. There are a few ways to represent these kinds of interactions which gives a large toolbox of ways to approach the problem.

The longer I work with a database the more I learn the dark corners of the dataset. Make sure you exclude the rows created by the test accounts listed in another table. Don't use the create_date field, use the real_create_date_v2 instead, unless it's not there, then just use create_date. Make sure you only get data from the latest snapshot for the key. Very quickly I end up with complex spaghetti SQL, which either contains monstrous subqueries or a chain of CREATE TEMPORARY TABLE.

The Jaccard Index is a useful measure of similarity between two sets. It makes sense for any two sets, is efficient to compute at scale and it's arithmetic complement is a metric. However for clustering it has one major disadvantage; small sets are never close to large sets. Suppose you have sets that you want to cluster together for analysis. For example each set could be a website and the elements are people who visit that website.

Two similar documents are likely to have many similar phrases relative to the number of words in the document. In particular if you're concerned with plagiarism and copyright, getting the same data through multiple sources, or finding versions of the same document this approach could be useful. In particular MinHash can quickly find pairs of items with a high Jaccard index, which we can run on sequences of w tokens. A hard question is what's the right number for w?

When comparing two texts it's useful to have a side-by-side comparison highlighting the differences. This is straightforward using HTML in Jupyter Notebooks with Python, and the inbuilt DiffLib. I used this to display job ads duplicated between different sites. For a long document it's important to align the sentences (otherwise it's hard to compare the differences), and highlight the individual differences at a word level. Overall the problems are breaking up a text into sentences and words, aligning the sentences, finding word level differences and displaying them side-by-side.

I created a simple custom recipe to show diffs between two texts in Prodigy. I intend to use this to annotate near-duplicates. The process was pretty easy, but I got tripped up a little. I've been extracting job titles and skills from the job ads in the Adzuna Job Salary Predictions Kaggle Competition. One thing I noticed is there are a lot of job ads that are almost exactly the se; sometimes between the train and test set which is a data leak.

Sequences of words are useful for characterising text and for understanding text. If two texts have many similar sequences of 6 or 7 words it's very likely they have a similar origin. When splitting apart text it can be useful to keep common phrases like "New York" together rather than treating them as the separate words "New" and "York". To do this we need a way of extracting and counting sequences of words.

I've been trying to use Named Entity Recogniser (NER) to extract job titles from the titles of job ads to better understand a collection of job ads. While NER is great, it's not the right tool for this job, and I'm going to switch to a counting based approach. NER models try to extract things like the names of people, places or products. SpaCy's NER model which I used is optimised to these cases (looking at things like capitalisation of words).

Over the past decade deep neural networks have revolutionised dealing with unstructured data. Problems like identifying what objects are in a video through generating realistic text to translating speech between languages that were intractable are now used in real-time production systems. You might think that today all problems on text, audio an images should be solved by training end-to-end neural networks. However rules and pipelines are still extremely valuable in building systems, and can leverage the information extracted from the black-box neural networks.

In a couple of hours I trained a reasonable job title Named Entity Recogniser for job ad titles using Prodigy, with over 70% accuracy. While 70% doesn't sound great it's a bit ambiguous what a job title is, and getting exactly the bounds of the job title can be a hard problem. It's definitely good enough to be useful, and could be improved. After thinking through an annotation scheme for job titles I wanted to try annotating and training a model.

When doing Named Entity Recognition it's important to think about how to set up the problem. There's a balance between what you're trying to achieve and what the algorithm can do easily. Coming up with an annotation scheme is hard, because as soon as you start annotating you notice lots of edge cases. This post will go through an example with extracting job titles from job ads. In our previous post we looked at what was in a job ad title and a way of extracting some common job titles from the ads.

There are many great tools for filtering, transforming and aggregating data like SQL, R dplyr and Python Pandas (not to mention Excel). But sometimes when I'm working on a remote server I want to quickly extract some information from a file without switching to one of these environments. The standard UNIX tools like uniq, sort, sed and awk can do blazing fast transformations on text files that don't fit in memory and are easy to chain together.