Machine Learning in SEM: Why you should hurry up
Machine Learning is nothing less than one of the biggest disruptions in working-history. Without any doubts. And it already affects a lot of fields in Search Engine Marketing. Even though we’re just at the beginning. If you understand how to use Machine Learning for SEM, you’ll basically understand how Google, Bing, Yandex and Co work and get a tremendously powerful competetive edge and a game-changing tool to push your campaign-efficiency.
How Machine Learning boost your Search Engine Marketing
Everything we do in Search Engine Marketing is based on data. Date we’ve collected from search engines and our websites.
The more data we get the harder it gets to make an impactful business decision manually. We drown in information.
And here Machine Learning comes in. It’s not just your lifebelt to keep swimming in the Big Data ocean. Machine Learning will become the speedboat that accelerates your SEM-strategy and execution.
It provides you the information you need to beat your competitor and to win the market. As soon as you enrich your SEM-skillset with Machine Learning knowledge your campaigns will become more precise, your spendings more efficient and your business decisions more sustainable.
Common fields and applications of Machine Learning in Search Engine Marketing are:
- Content personalization
- bidding automation
- Data analytics
- Market- & audience segmentation
- Automation of tasks and reportings
- Content Creation
- Attribution modelling
Understand what Google Ads is doing in the background
Smartbidding, Responsive Search Ads and In-Market-Audiences are powerful tools. But for Search Campaign Managers, like you and me, they’re black boxes. We don’t know what’s going on behind the curtain and need to rely on the outcome Google provides. We have no possibilities to change the learning rate of the algorithm or delete features we don’t want to include.
There is no chance to intervene in what is happening, but what exactly happens is not such a secret as many would expect. It might surprise you, but a lot of stuff the Google teams are working on was and is published. the knowledge is available for you and me. One mission of this blog is to provide this knowledge to help Campaign Managers gaining knowledge about the Machine Learning techniques behind the tools they use.
More information about how Google uses Machine Learning in ad tech you’ll find here: https://sem-smartation.com/calculation-of-probabilities-how-machine-learning-rules-ad-tech/
What is Machine Learning exactly? Let’s dive a bit deeper
Machine Learning is a subfield of Artificial Intelligence. In a very broad definition AI can be seen as the simulation of human intelligence processes. Therefore it combines several subfields and different levels of machine capability.
If you imagine it as a circle you’ll find ‘Search and Planning’ on the outer edge. Then ‘Reasoning & Knowledge’, followed by ‘Perception’, the ability to move objects and the language interaction between humans and computer-programs called ‘Natural Language Processing’.
And finally you’ll see ‘Learning’ in the innermost circle. Within Learning you can find „Deep Learning“, the very final stage. The (hopefully not) endgame: machines that are able to make decisions and build conclusions.
Following Arthur Samuel’s much quoted definition, Machine Learning is the „field of study that gives computers the ability to learn without being explicitly programmed”. From Tom Mitchel’s explanation we can extract, that the learning experience the machine makes, improves it’s performance on a certain task.
With other words: a machine develops knowledge by itself based on data and is able to solve problems in a specific field. The machine grows with the challenge and gets „smart“. The more data, the more experience. The more experience, the better/smarter the machine.
Most of Machine Learning done by today is Supervised Learning
In Machine Learning we distinguish three different kind of Learning problems
- Supervised Learning: working with labeled data
I’d rougly assume that 75% of applied Machine Learning done by today happens in Supervised Machine Learning.
If we use Supervised Learning, we train an algorithm by providing a labeled data-input (independent variables) by ourselves. The crucial point is, that WE train the algorithm. We’re the supervisors and teachers. We already know about the output (the dependent labeled value, e.g. salesnumbers).
So the algorithms task is to find the most efficient way to reach this specific outcome-number by providing predictions of correlations between different datapoints (linear regression) and probabilities of belongings to specific groups of data (logistic regression).
To handle the massiv amount of data companies like Google use a model-type called „Neural Network“, where we have a huge and closely linked net of algorithm-layers, which get updated by multiplying them with weight-parameters, which determine their connection.
It’s just a logic extension that Deep Neural Networks followed on the era of Big Data.
And it’s not a coincidence that Google released it’s Machine Learning ad-features shortly after the company acquired the UK-based AI-start-up Deep-Mind.
2. Unsupervised Learning: working with unlabeled data
Unlike Supervised Learning, Unsupervised Learning does not come along with labeled data. We don’t split between dependent and independent variables here. Therefore the result is unknown in the beginning.
Imagine a big bunch of data. What an Unsupervised Learning algorithm does, is to cluster and group this dataset based on similarities, patterns and associations. Therefore this technique is very popular and mostly common in market segmentation and audience clustering.
An implementation guide for audience-segmentation in Python you’ll find here: https://sem-smartation.com/audience-clustering/
3. Reinforcement Learning: learning from environment
In contrast to Supervised Learning, Reinforcement Learning is still waiting for it’s breakthrough in Applied-Machine-Learning. Today you’ll find it mainly in video-games and robotics.
Reinforcement models are following a semi-supervised approached by learning from their own environment. Imagine a robot which is supposed to climb up a tree. For the first 50 trials it will fail and fall down after a few meters. But eventually it starts to understand which branches give a strong hold. It makes experience and improves performance. The robot climbs higher and higher. It learned from try and error.
The Machine Learning Process: Predict – Err – Learn
The Machine Learning process itself is always the same. The scheme you’ll find everywhere in Machine Learning is: „Predict – Err – Learn“.
Based on a huge amount of data an algorithm makes a prediction. We then calculate the error-distance of our calculation. With respect to a reference number the algorithm will calculate how good or how bad that prediction actually was.
This is done by the Error-Function (also known as cost function), which is used to describe diffusion. Based on this calculation the machine will change it’s coefficients or weights, by updating the paramaters, to get a better result in the next trial. This process runs by the model until the perfect value (the global minimum) is reached.
If this whole process sounds familiar, you’re right. It’s quite similar to the common approach of a marketing analysis (research– strategy – implementation – analysis – optimization). We prepare, we test, we fail, we step back, based on new knowledge we learn and try again. The same does your algorithm. But much faster and with millions of different datapoints.
More information about the steps in the Machine Learning process you’ll find here: https://sem-smartation.com/the-four-steps-of-machine-learning/