Fast Data Science

How can you use data science and machine learning to predict how much your customers will spend?

Machine learning model to predict turnover of employees in an organisation of 3 million workers

I have worked on a number of different projects where a client needed to parse scientific literature and identify occurrences of molecules or proteins.

As an example, the molecule on the right is Aspirin. This is still a trademark of Bayer in some countries. But in a paper it could appear under acetylsalicylic acid, 2-acetoxybenzenecarboxylic acid,
C9H8O4, or a number of identifiers such as DB00945. There could also be identifiers that refer to other molecules, or identifiers that refer to only one version of a molecule.

Another example I have encountered often in clinical papers is the gene ERBB2, which is important in certain types of breast cancer. ERBB2 is also called Erb-B2 Receptor Tyrosine Kinase, HER2, HER-2 and many other names. These names often also refer to the protein expressed by the gene. Many names are similar to common English words, and are not always capitalised in text.

Because of these pathological effects, the task of identifying names of proteins, genes and molecules in scientific literature is fraught with difficulty.

I have developed several tried and tested techniques to disambiguate these terms. Usually I need a number of annotated examples to start with, and I will train a machine learning model to learn from these examples and annotate new publications as they come in.

This can be deployed on the client’s servers and provide daily updates on a dashboard. This allows a client to monitor the literature in real time for publications around a particular molecule, protein or gene, or to spot trends in advance.

A client in the retail industry had a fleet of vehicles delivering produce at different times of day. They used third party logistics software to plan the delivery schedules, however an element of the delivery schedules that was hard to plan was the unloading time of the vehicle when it arrived at the store.

Fortunately there was a system in place for recording vehicle ignition events, GPS location, and geofencing to identify the arrival and departure times of delivery vehicles, and past schedules were available to identify the quantity and type of product delivered on each drop, which driver was in charge, and the time of day and type of vehicle used.

Using this trove of logged data I was able to train a simple regression model that would predict the unloading time of any future delivery at the time that the schedule is being generated.

This allowed the client to save money on driver overtime, disruption caused by late deliveries, and fines due to drivers working longer than their legally permitted hours.

With one customer in the recruitment industry I found that the web form that was used for jobseeker signup was very long. By analysing the fields in the form I was able to establish that users were confused by some fields, and lingered for a long time in some areas.

Since users also uploaded their CV which contains explicitly lots of personal information, as well as implicit information such as the job type or salary that someone was looking for, I was able to train a deep neural network on past signup data over several years, to analyse the CV and fill out some of the fields in the signup form automatically. 

This allowed a field to be removed, which boosted the conversion rate of the form by 7%, measured by A/B testing.

This is one small example of what can be achieved combining text analysis and conversion optimisation techniques.

If you have a long signup form on your website please let me know and I may be able to deploy a machine learning model to improve the user experience and boost your conversions.

When a pharmaceutical company develops a drug, it needs to pass through several phases of trials before it can be approved by regulators.

Before the trial is run, the drug developer writes a document called a protocol. This contains key information about how long the trial will run for, what is the risk to participants, what kind of treatment is being investigated, etc.

The problem is that each protocol is up to 200 pages long and the structure can vary.

For one pharma company I developed and trained a deep learning tool to predict more than 50 output variables from a clinical trial protocol. This allows pharma companies and regulators to analyse and quantify large numbers of protocols, allowing more accurate cost estimation.

The technique can be extended to other industries where large unstructured or semi-structured documents are the norm.

If you have a problem of this nature please get in contact with me and I will be glad to discuss.