Owned Machine Learning: Why company-owned algorithms are vital in SEA & online advertising


  Why company-owned algorithms are vital in SEA & online advertising for Machine Learning



Work in online marketing without tools such as Automated Bidding, Affinity Targeting or Responsive Search Ads. Hard to imagine, isn't it? They are all an integral part of our everyday office life. And they are all based on machine learning. The problem here: As account managers, we often sit in front of a black box. Valuable knowledge about the user, his behavior and interest remain on the servers of Google and Facebook. In-market audiences, for example, are provided as a finished marketing segment. The backgrounds and calculations of these mappings remain hidden. It is precisely this knowledge that determines long-term success and failure in marketing. With Owned Machine Learning, self-implemented and trained algorithms, companies can quickly draw this knowledge from large amounts of data and make use of it. This article is intended to show how machine learning works in principle, why it is necessary for the long-term success of the company and how we can integrate this tool meaningfully into our online marketing with help wanted .
This article focuses on the use of machine learning in SEA, Facebook advertising and Paid Advertising. Also, the article Importance of Machine Learning, AI & Rankbrain for SEO & Google with help wanted

What is Machine Learning anyway?
"What we want is a machine that can learn from experience"- Alan Turing (mathematician, IT pioneer)

Help wanted Machine learning is one of many sub-areas of artificial intelligence. According to the common definition of the American computer scientist Tom Mitchell, the basic idea of machine learning is that a computer program automatically improves its performance through newly established experiences (data) in a particular area. The advantage of this is that the program does not have to be permanently reprogrammed with thousands of lines of code. A wide range of mathematical algorithms ensure automatic processing of data and thus the learning process with help wanted 

Probably the best-known examples of machine learning implementations are the referral systems used by Amazon and Netflix. Online shopping and video streaming as we know it today would not be possible without algorithms and the processing of user data.





Even self-driving cars or social media feeds would not exist without artificial intelligence. Currently, machine learning may save many lives in the wake of the Corona epidemic, as scientists and virologists can use the algorithms to make predictions about the spread.

Without machine learning, Amazon's recommendations wouldn't be as accurate and effective

Due to the interplay of big data and an ever-increasing level of CPU computing power, it has become increasingly possible in recent decades, of course, also in online marketing, to train models with data and let them "learn". Algorithms are now an integral part of most marketers' day-to-day digital business. Google and Facebook have been instrumental in driving this development. Both account for more than 85% of their sales from online advertising and have made great progress in the field of deep learning implementation in the 2010s.

A much-publicized article about Google's landmark decision to organize the entire company around artificial intelligence and machine learning was Gideon Lewis-Kraus' November 2016 NY Times feature: https://www.nytimes.com/2016/12/14/magazine/the-great-ai-awakening.html

The Great A.I. Awakening, Source: New York Times

Machine Learning Potentials in Online Marketing
"Accuracy is, in every case, advantageous to beauty"- David Hume (philosopher)

Unfortunately, we humans are generally very bad at recognizing patterns in large amounts of data (such as marketing reports). It becomes even more problematic if conclusions and learnings for future action are to be drawn from these data sets. Each of us has an individual wealth of experience that shapes our ideas, values and worldview. It is not always possible to eliminate this "cognitive bias" during an analysis. Based on my marketing experience, if I assume that the male 50+ audience tends to respond more positively to advertising a new sports car, I will probably recognize these patterns in the campaign's analysis record and ignore other – perhaps much more relevant facts.

We don't want to talk about distractions in the office, of tiredness and of hunger, which leads us, people, to make mistakes.

A machine learning model does not know all these sensitivities, so the benefits of ML applications in online marketing are obvious. By reducing human bias and bias, business-relevant decisions are increasingly made based on data.

Also, processes can be enormously accelerated by the computing power of computers and digital advertising measures, such as campaigns in search advertising and search engine advertising, can be automated. The result is much greater effectiveness and a reduction in marketing costs.

Of course, machine learning is not a miracle cure in online marketing either. Rather, it should be embedded and complemented as a potential – but extremely powerful – tool in the existing strategy and marketing measures.

And machine learning is not always the optimal solution for a marketing problem. For example, if there is too little data that the algorithm could process. Here, a balance must be taken between different approaches to solutions. The saying "If you only have a hammer as a tool, you see a nail in every problem" describes this quite well.

Catherine Williams, Chief Data Scientist at AT &T, gave a compact overview of the myths and realities of machine learning in digital marketing last summer:

Monitored vs. Unsupervised Machine Learning
In the field of machine learning, there are three different categories:

In supervised learning, conclusions and predictions about future events are made based on structured, historical input and result data (e.g. user data and conversions). The output value always acts as a dependent "label" in correlation to the independent input variables.
It is estimated that about 70% of all machine learning applications are still found in this well-researched field. Typical Supervised ML applications in marketing are, for example, the data-driven revenue forecast or the division of customer groups into predefined segments and classes (classification).

Two typical 'supervised' machine learning problems: regression and classification

On the other hand, there is unsupervised (unsupervised) learning. These algorithms and models are used for pattern recognition in large amounts of data without predefined dependency or correlation. In online marketing, for example, this is useful for accurately dividing website visitors into different as yet undefined groups, and then addressing them with a tailored advertising message (clustering).

Unsupervised algorithms reliably detect patterns in large amounts of data

In strengthening (reinforcement) learning, the learning process of the machine takes place by interaction with a generally predefined environment. Positive behavior is rewarded and thus encourages the system to act in this way again in the future. Many of these applications and developments are currently in the areas of gaming and robotics.
The challenge in supervised learning is above all in data preprocessing and to ask the right questions in advance, which the system should learn. The model can only solve the tasks and problems that we teach it and understand only the correlations that we predefine. Accordingly, human bias plays a greater role here than in unsupervised learning. A problem currently being addressed by teams at Google, among others
Even before the learning process, it is possible to ensure that the model requires less computing capacity and thus delivers solid results faster. Adjusting screws here are, for example, the removal of unnecessary columns (such as the currency column in keyword reports) and the label encoding of strings (text modules, such as keywords or locations), i.e. the conversion to numerical values, which can be interpreted more easily by the system.

By visualizing the data before the actual training, for example using histograms or scatter diagrams, we can get an overview in advance of whether we have enough data in the set from which the machine can learn. In this way, we can steer the quality of the significance of our model in the right direction at an early stage and avoid phenomena such as overfitting and underfitting.

Histograms can be used to create the first rough hypotheses about the model

How a machine learns: Mistakes explicitly desired
"The beautiful thing about learning is nobody can take it away from you"- B.B. King (Blues legend)

In the 1930s, the Swiss psychologist Jean Piaget made an interesting discovery among his offspring. It is true that the children, who are still very young, made rapid learning and sensory progress when they interacted with their environment. They applied the knowledge of experience to new situations to grasp them.
However, their already solidified first conceptions of the world were at first only very slowly changed or replaced by new information. "Updated" we would say today.

For example, if one of the children was already familiar with the scheme of a dog and then met a cat for the first time, it was also a dog in its eyes, since this animal runs on four legs and has a fur. Only additional experiences due to more frequent encounters with the cat led to the child being able to distinguish the cat from the dog according to various criteria.

Piaget referred to this as assimilation (knowledge is absorbed by experience) and accommodation (knowledge is changed and replaced).

Assimilation and failed accommodation in humans

Interestingly, the machine learning process is not so different from what we go through as children. Comparisons of untrained machines with young children appear quite frequently in scientific publications.
What both have in common is a relatively manageable to non-existent treasure trove of experience. Data on which conclusions could be drawn are initially sparse or non-existent.

Thus, even a machine with a manageable amount of data would first classify cats as dogs with great confidence. Only that the machine sees the animals as columns and rows of data and concludes calculation and algorithms. With increasing experience, through further data intake and additional information about cats, the ability to reliably distinguish the two animals from each other would greatly improve.

The learning process of a machine learning model always goes through four steps. This process is found in all ML models and applications; from social media feed to self-driving cars.

Data acquisition/hypothesis (assimilation)
Error Calculation
Parameter Updates/Error Minimization (Accommodation)
Evaluation
First, a hypothesis is calculated based on the available data, which should describe the existing observations as well as possible. In most cases, there are deviations, which are calculated in the second step with the help of an error function.

Hypothesis and error calculation (red lines) in the linear model

The actual learning is then carried out in step number three, in which the calculated errors are minimized until an optimal value is reached. A fourth step is limited to the area of supervised learning, where we already know the output. By splitting into a training and test data set, we can evaluate our model here with the help of previously unknown data.

A detailed description of the machine learning process can be provided by Google's Machine Learning Crash course: https://developers.google.com/machine-learning/crash-course/reducing-loss/an-iterative-approach

What can be inferred from this process is the conclusion that a machine learning model is more accurate and accurate the more experience it has. In other words, the more training data available, the more accurate the prediction becomes. Gigantic amounts of data with positive and negative data are needed. This is the great competitive advantage of tech companies such as Google and Facebook. Since it cannot be assumed that they will eventually share their knowledge and data with us (even if it is insisted on by politicians), we as a company must take action ourselves.
Why Owned Machine Learning is existential in online marketing today
"Marketing aims to know and understand the customer so well the product or service fits him and sells itself"- Peter Drucker (economist)

The machine learning tools from Google and Facebook have fundamentally revolutionized our work in online marketing. Today, advertising messages can be played to the right user in the right place faster and more effectively than ever before. However, the price we as marketers pay for the effectiveness gained is an increased degree of dependence and lack of transparency. We push data into smart bidding black boxes and hope for the best. Often the result is satisfactory, so that sufficient conversions can be recorded in the next reporting.

How these came about in the end, however, is eludes our perspective. But that is where the real value lies. If we want to be successful as a company in the long term, we need exactly this information. Target group knowledge is known to be a power in marketing. Who knows which trigger is most effective for which target group, can effectively focus their resources on it.

For example, Google has been offering a so-called "Bid Strategy Report" for some time, but also mentions that the "Top Signals" shown are only examples. There is no information on the weighting of the variables among themselves. For example, we do not see how important the user signal mobile device has to the conversion probability compared to a certain day of the week or a particular interest.

Bid-Strategy report in Google Ads, Source: searchengineland.com

It feels a bit like driving a self-driving car that doesn't inform us of the planned route. We sit back and no longer care about storing knowledge about the optimal pathway in our brains.
But one day the autopilot fails and we have to get back on the wheel. The road layout has changed in the meantime. What now? We no longer know the fastest and most effective route to the destination.

What if Google or Facebook pull (or get pulled) the day after tomorrow.

Could we then say which variables have the greatest leverage effect on our most valuable target group?
Could we still say, in a well-founded and data-driven manner, which target group is the most valuable for us?
This is where it is worthwhile to start and use its machine learning models to identify high-turnover target groups and leverage in the run-up to a campaign launch.
Our biggest asset as an account manager is unsurprisingly what Google & Co craves the most.
First-party data shouldn't lie idle
"More data beats clever algorithms, but better data beats more data "- Peter Norvig (Director of Research, Google)

The real power in machine learning lies in the data. Who has the most data always wins, right? Not quite.

The quality of the data is also crucial. It's no surprise that CRM integration tools like 'Customer Match' are so heavily touted by the big tech companies. The ad companies know about the relevance and preciousness of direct customer data.

After all, what we want in online marketing are new users who are as similar as possible to our actual existing customers. Behaviour, interests, demographics. The hypothesis is that those lookalike users tend to buy to the same extent as their targeting mirror images.

What we are looking for in online marketing are statistical twins of our customers

We want to find these users through statistical models in machine learning. We want to identify patterns and put the puzzle together by recognizing a potential customer like that. If User X has the same characteristics as existing customer Y, we are more likely to be able to convince this user to buy our product.

Only First-Party data offers us this information value. Data from users who have already interacted with their brand. Collected, for example, through website tracking, email subscribers, app engagements, or digital checkout systems.

From a machine learning perspective, as described, it is a matter of using data as input that reflects our hypothesis as well as possible. However, the further away the data is from the customer (third party data), the more our hypothesis is blurred. The significance of the model becomes less precise. It becomes unusable to make business decisions based on it.

"The answers are all out there. We just need to ask the right questions" - Oscar Wilde (writer)

At the end of the day, it's about being able to ask the right questions with the available data. We ask for the value of individual variables and information and conclude the actions of new user groups based on conversions that have already been made.

Thus, Owned Machine Learning pursues the approach of building and training our effective models, which we can use flexibly at any time in combination with our data. This will enable us to make better, faster and most informed decisions. Also, target groups and projects can be prioritized more specifically.

Online Marketing Implementation: Owned Machine Learning in Use
"The amount of work we can automate with AI is vastly bigger than before"- Andrew Ng (Founder Google Brain Project)

Where can Owned Machine Learning in marketing make our daily work easier?

We can outsource numerous tasks to our data models, which can effectively take over these tasks with some training and high-quality data.

Audience Prioritization: Unfortunately, we can't create a personalized user experience for every single audience. We always lack the time and often the money for this. But we don't have to. Because if we use machine learning to find out which target groups are most likely to convert, we can focus our limited resources on them.

Affinity Audiences in Google Ads, Which is the most valuable for me?

Customer classification: Is a user one of our more valuable target groups? Is there a high probability of a high Customer Lifetime Value? or costs and effort are not worthwhile to use advertising. We can automate this classification with machine learning. A detailed description of the approach, available from Google itself: https://cloud.google.com/solutions/machine-learning/clv-prediction-with-offline-training-intro
Keyword segmentation: How does our target group search? Are there any search terms and keywords that we can assign to the Lower Funnel? (More about Customer Journey and Sales Funnel) Classification algorithms can help us here. A way of implementing with the Logistic Regression algorithm can be found here: https://sem-smartation.com/how-to-build-a-keyword-classifier-with-machine-learning/
Audience Clustering: Where does it make sense to set the cut and separate two or more user groups. Let's let the data decide. For problems of this kind, the k-means algorithm is often used: http://sem-smartation.com/audience-clustering/

Audience segmentation with machine learning

Marketing Channel Attribution: We work in parallel with multiple channels with Machine Learning, which allows us to determine the value of individual channels in the marketing mix for specific user groups. Further information on the implementation can be found here: https://towardsdatascience.com/data-driven-marketing-attribution-1a28d2e613a0
Forecasting and budget calculation: How many conversions can we expect for Budget X? How many clicks do we get with Budget Y? Every marketer knows this problem. A classic case for logistic regression and time series algorithms.
Anomaly detection in reporting: Reporting and reporting are among the biggest time-eaters in online marketing. Let your machine learning model scour your data and effectively detect outliers or potential trends. One of the most detailed explanations of Anomaly Detection Algorithms can be found in Andrew Ng's Machine Learning Course at coursera.org

If we want to be successful as a company in the long term, we need to know how our most valuable target groups think and feel. There is no way around it.

With Owned Machine Learning, specially trained models, we can efficiently filter out these findings from large amounts of data. We can then use this knowledge to provide the incredibly strong – but opaque – machine learning models of Google, Facebook & Co with corresponding clearer signals as input. For example, by focusing on in-market or affinity audiences that are particularly valuable to us.

The implementation should be strongly based on the available data and the respective problem or question. We, online marketers, have countless algorithms at our disposal, each one has advantages and disadvantages.

Owned Machine Learning can therefore be focused on two core aspects. We gain long-term knowledge about the behavior of our target group and integrate high-quality data into our self-developed models to make our online marketing more effective.

There is no predetermined path. Even if it may be a journey with a predetermined, clearly defined goal: on the way, we will learn things about our target group through machine power that we would not have expected.

Artificial intelligence and machine learning are still at the beginning of their development. Even in online marketing, we haven't seen everything yet, and the untapped potential is gigantic.
If we deal with this topic at an early stage, we will secure a competitive advantage and, in the long term, the customer's favor.
seo wiki,seo tools,seo tutorial,seo services,what is seo marketing,seo course,seo google,how to do seo,on-page seo,what is seo in digital marketing,


4 Comments

Thanks for your interest

  1. This comment has been removed by the author.

    ReplyDelete
  2. Very Informative and creative contents. This concept is a good way to enhance the knowledge. thanks for sharing. Continue to share your knowledge through articles like these, and keep posting on

    Data Engineering Services 

    Artificial Intelligence Solutions

    Data Analytics Services

    Business Intelligence Solutions

    ReplyDelete
  3. Thanks for sharing such beautiful information with us. Please keep sharing!

    Industry’s Top Monetization Platform for Top Publishers

    ReplyDelete
  4. I appreciate you taking the time and effort to share your knowledge. This material proved to be really efficient and beneficial to me. Thank you very much for providing this information. Continue to write your blog.

    Data Engineering Services 

    Artificial Intelligence Services

    Data Analytics Services

    Data Modernization Services

    ReplyDelete
Previous Post Next Post