Databowl has announced the launch of its Machine Learning Department - SKUNKWORX.
Skunkworx was specifically developed to further the capabilities of Databowl’s in-depth data analysis and stringent data validation measures. Databowl CEO Simon Delaney says “At Databowl we pride ourselves on having the most powerful software in B2C Marketing. We have the best developers in the world working here and the creation of Skunkworx reaffirms our position as market leaders.”
So far Skunkworx has only been used by an exclusive group of Databowl’s Blue Chip clients, but now the option is available for everyone to begin using it.
We spoke to Head of Skunkworx Malcolm Kring to learn more about the benefits and potential of using AI in Performance Marketing.
Firstly, why was Skunkworx created?
Malcolm Kring: Skunkworx was created because we saw an opportunity to add tons of value to what Databowl offers. Our customers have all this validated, formatted, high-quality data just sitting in the system and we realised could put it to work for them in all sorts of productive ways. Personally, I find this field and its potential very exciting and despite what we’ve accomplished already I think we’re just scratching the surface of what’s possible with machine learning with Databowl.
What exactly is AI? And what’s the difference between that and machine learning?
MK: In marketing, AI is a buzzword. It can refer to pretty much anything because the term ‘intelligence’ is so broad. Flies are ‘intelligent’ because they’re able to avoid predators and swung newspapers, but that doesn’t mean that they’re smart.
A lot of companies out there are using ‘AI' to refer to simple if-then conditions or other decision/scoring systems set up by humans. This is really misleading and, in my opinion, abusing the term. Skunkworx was set up as an AI Department in a broad sense, but currently, we specialise in machine learning.
Machine learning has a much narrower and more precise meaning, though it still covers a large number of ideas and techniques. Machine learning is basically setting a computer up to teach itself. This is especially useful for data analysis because there can be lots of complicated relationships between variables and lots of small and subtle effects. Most machine learning algorithms try lots of combinations really quickly to form a good approximation of the data. Once you have a way to approximate the data (a model), you can put new data through that model and get an accurate prediction. This is all done without any human input.
What types of companies would benefit from machine learning? How can it be used in performance marketing?
MK: Any B2C company can benefit from what we’re currently offering. An algorithm is trained on data and produces a model of your data which can recognise leads with potential and leads that are unlikely to convert. This is expressed as a score between 0-100 that gets attached to the lead as soon as it is received. What’s cool is that the predictions automatically get better as the algorithm receives more data to learn from!
Performance marketers can use this score to route the leads to different salespeople, put more resources into leads with lots of potential, and even reject leads with very little potential. The algorithms are very good at recognising leads that have virtually no potential, so it can be a real time saver and you can focus on more promising leads.
How do you make a model work?
MK: We’ve worked really hard on a slick UX, but there is a lot going on under the hood. Each model created is totally unique and bespoke to that customer’s data. We run the data through multiple different algorithms and basically have a competition between the algorithms to see which one produces the best model, which we then use.
There are a number of other more technical processes that we use such as hyper-parameter optimisation, resampling, feature selection, and others. These all happen automatically and ensure that customers have the most predictive and robust model possible.
Again, we want to make sure that the user has the best results, but insulate them from the technical stuff so that they can focus on their creative output, which machines can’t do.
What is the difference between human-built models and machine learning?
MK: Quality. Humans are fantastic at a lot of things, but in-depth data analysis is not one of them. Machines are able to analyse data at a rate, scale, and with a precision that we can’t even begin to match.
Problems are best predicted by looking at the sum of many small effects. These patterns are difficult for humans to detect. Compared to algorithms, systems set up by humans are rough and suffer from bias, incompleteness, and frankly, error. This can lead to terrible predictions, which can be worse than no prediction at all.
Machines are naive in the best way. They are able to look at data with completely fresh eyes and analyse each dataset on it own merits. This helps to find the factors that are predictive for that precise dataset. What factors are predictive can change dramatically between datasets so the flexibility, depth, and self-contained nature of these algorithms really shines.
What other features does Skunkworx have? (And how can people use them?)
MK: We’re really proud of our analytics features that we’ve been building. Once we built the predictive algorithms, we realised that customers want more than just an accurate prediction, they want to see how the prediction is made so they can understand their campaign.
We’ve built a set of tools to visually explore the data in the same way a machine learning algorithm would. This lets you compare leads that you deem ‘successful’ and ‘unsuccessful’ across a large number of metrics. The visual elements really give you information at a glance and you can explore which features are over and underrepresented in the successful data. You can even compare the way features interact to get a really precise understanding of your data. This is a powerful tool for gaining insight into your data.
Another thing that machine learning does is fraud detection. Most of the fraudulent data is caught before it gets to us by other parts of Databowl, but a little can slip through. However, because this fraudulent data is created by a machine, it has distinctive patterns and can be caught using a better machine.
We’ve seen some really impressive results in the lead scoring product where anything it scores under 15 (out of 100) is virtually guaranteed to never convert. This means you can safely discard or reject those leads and save time and money.
When can people start using Skunkworx?
MK: Skunkworx is fully featured and ready to go! We’ve got lead prediction and analytics launched and in use. We’re also really excited about the new features in development.
There is so much potential for machine learning in Databowl and we know our customers are going to keep getting more value from it as time goes on. You can request a Databowl Skunkworx Demo here.