Among the many buzzwords circulating in the world of digital analytics - first-party data, server-side tracking or cookieless tracking, to name but a few - the term machine learning is also being used with increasing frequency. All of these terms have their raison d'être in digital analysis, provided they are used with the necessary background knowledge and in the right context.

This article is intended to help you categorise the term machine learning correctly in the context of digital analysis. We will briefly and concisely explain what machine learning is (and what it is not), how and where it can be used and what needs to be taken into particular consideration.

What is machine learning

Machine learning is a huge field within artificial intelligence, which has many branches and consists of countless specialised areas. Describing machine learning as a whole in this short article is therefore impossible and could not do justice to the diversity of this field.

Broadly speaking, however, the aim of machine learning is to automatically calculate statistical models from typically very extensive data sets. These models can then be used to analyse known data and thus reveal correlations or regularities that are hidden from the human eye. In addition, future, as yet unknown data can be categorised according to the same pattern and thus forecasts can be calculated in real time.

What machine learning is not

Unfortunately, the term machine learning is often used in advertising, giving the impression that machine learning is a miracle cure for all problems. Just because a product is labelled with the word machine learning, the quality of this product is not automatically better than the competition. The quality of a product should always be measured by what it does, not how it does it.

It is often forgotten that machine learning is not the same as "machine thinking"! Machine learning is a tool. The use of a tool must be learnt. A tool must be used for what it was invented for. A tool does not provide the blueprint, it must be guided. This means that machine learning can neither do the thinking for us, nor the conceptualisation or planning of a project. Machine learning can only be used successfully if the goal, the associated factors, the context and the environment - i.e. the technical use case - are given.

Prerequisites

In order for machine learning to be used optimally and to realise its full potential, a number of basic requirements must be met:

Data. As the basic idea of machine learning is that a model is trained using data, data quality is the most important prerequisite. How well the model learns depends on how good the data is that is fed into the system. Or to put it another way: "Garbage in, garbage out."

Data pipeline. Goals that are to be achieved with machine learning are often linked to one or more systems from which data is generated. It is essential that the data pipeline from the source system to the model and back supports the objective being pursued. This applies in particular if the model provides for real-time data processing.

Data preparation. In order for the model to be able to process the data, it must be supplied to the model in an appropriately prepared form. It is advantageous here if the data pipeline can route the data as close as possible to the model and if initial preparation or structuring of the data has already been carried out in the pipeline.

To fulfil these requirements, specialised products such as those from Tealium iQ, which cover the pipeline from the website to the model in real time, are particularly suitable in the digital analytics environment.and the associated structured data layer and, in between, also take over data preparation via the Customer Data Hub, where the data can already be aggregated and enriched at session and user level.

As a fully integrated system, Tealium Predict in particular offers an easy entry into the world of machine learning.

Possible use cases in digital analysis

Scoring

Scoring is about training the model with behavioural data (e.g. sessions) that indicate whether the behaviour was successful or not. Success can be defined, for example, as a completed purchase or other desired behaviour on the part of the visitor.

This teaches the model to predict how close this behaviour is to success when a new (partial) behaviour occurs. This is indicated in the form of a score, for example a probability. This information can then be used to either guide the visitor directly towards success in real time with a personalised website or to target them again afterwards. The respective measure then depends on the score determined by the model.

Attribution

During attribution, the model can be trained with information on individual touchpoints (e.g. access source from which the website was visited). The aim of the model in this case is to learn which access sources occurred in the visitor's overall journey and to what extent these were - also indirectly - responsible for the success.

Affinity

Determining the affinity for a specific topic or product category can be another interesting use case with machine learning. In this case, the focus is on which areas of the website were visited by visitors who ultimately made a purchase.

The model can thus learn which topics are relevant for buyers of different product categories. This knowledge can then be used either to optimise the user experience or for active advertising, both online and offline.

Conclusion

Well planned, correctly implemented and, above all, used under ideal conditions, machine learning certainly has great potential in digital analysis and any company that wants to get more out of its data. In principle, machine learning does not require a particularly expensive, large or powerful infrastructure, but often fails due to a lack of agility on the part of the company (organisational or technical) to allow the data to flow and to acquire the necessary technologies.

Would you like to jump on the machine learning bandwagon with your company? We would be happy to advise you!