AI, machine learning, reinforcement learning… A journey into the heart of predictivemarketing

Written by Thibaut PIRIOU LE LAN on Jan 8, 2019 2:16:57 PM

Data exploitation that is even more targeted than Big Data is now available, thanks to Artificial Intelligence techniques such as Machine Learning or Reinforcement Learning. These constitute the heart of predictive marketing in the Advalo Platform. Meeting and decryption with Nicolas, a Data Scientist at Advalo.

 

AI_VISUEL_EN_visuel 

“These techniques transform mass marketing into a more acceptable personalized relationship which is more meaningful to consumers. A real relationship, much as the one a storekeeper would have had with his customers 50 years ago.”

Nicolas, Data Scientist at Advalo

 

Can you tell us more about your job?

A Data Scientist manages, analyses and exploits mass data. This work is comparable to the evolution of a statistician in the big data era. In practical terms, my job involves using state of the art standards to decide whether AI (Artificial Intelligence) and all the techniques relating to it can be used in specific marketing scenarios.

So there is a lot of R&D with adaptation of open source algorithms (resulting from research and sharing among developers) for creating exploitable models for our retail and automobile customers.  -When would someone who owns such a vehicle start thinking about upgrading to a new model?
-How can you determine whether someone is thinking about purchasing a new pair of jeans after surfing on the internet, his or her likes on social networks and even IRL (in real life) movements ...

AI allows us to provide an accurate response to these types of queries: working on moments, finding new insights. In addition to his or her expertise, a good Data Scientist must have a multi-disciplinary approach, which takes into account the constraints inherent in the domain he/she works in and be good at communicating the results to non- experts.

 

So AI is the new digital marketing buzz word, but what is it exactly?

To give you an all-encompassing definition, I would say that AI is a set of theories and techniques that are applied toward creating machines that are able to simulate human intelligence. Having said that, strong AI can be set apart from weak AI: Strong AI aims to resemble human cognitive abilities (much like in science fiction movies), which makes for a good objective for theoretical research, but remains a far away dream despite the considerable progress that has been made.

Conversely, weak AI is what we use in the operational sense, such as for digital marketing: this involves setting up ultra- specialized algorithms to perform a specific task and operate a device that uses Big Data autonomously.

 

Precisely, so what is Machine Learning?

Machine Learning is AI technology that allows computers to learn, without having to program them to do so. This is therefore a very broad set of algorithms which includes learning, which can go from a simple mathematical function with just a few parameters requiring optimization up to multi-layered neural networks (“Deep Learning”) with billions of parameters that can be calibrated by the machine as it learns.

For example, we can implement an algorithm that is able to estimate the fair value of an apartment based on different pieces of information which the machine itself will collect (area, age, equipment, municipality, neighborhood, distance from schools, transportation and shops, values of similar assets, etc.). In fact, Machine Learning covers three categories: supervised learning, unsupervised learning and reinforcement learning. Of course, in the digital marketing field, these different techniques complement or even mix with each other.

 

What exactly is Reinforcement Learning?

In this case, unlike Machine Learning (whether supervised or unsupervised), the dataset (associations between observations and variables) on which the algorithm will work is not initially given: it will be discovered as the agent interacts with his environment. I can give you an actual example with the Bandit Manchot algorithm which can be viewed as an extension of the classic A/B testing: instead of waiting for the end of the testing period to arrive at a conclusion, we can “exploit” (make the optimum decision as learned at a time T) and “explore” (test sub-optimal options in order to continue to learn) at the same time: during an email campaign it is possible to test several combinations of email content, objects, offers, etc. to arrive at the most effective mix.

 

How do these techniques function on the Advalo platform?

They are at the heart of the solutions we propose since they allow predictive models to be implemented. The first step is to reconcile offline and online data to create a "classic" DMP. We can then obtain an omni-channel vision using pure statistical techniques (Business Intelligence, Analytics, Datamining, etc.), but without an automated model, this is simply an "image" taken at a precise time.

With the contribution of AI, we can really take action in real time to optimize targeting, offers, campaigns and discover consumer insights: this is almost individualized, but above all automated marketing. It should be noted that these techniques can also be used in a less visible way, such as for the enrichment of our product repositories using NLP (Natural Language Processing - the ability of a program to understand human language) or Deep Learning (for example, with the implementation of product image recognition).

 

In actual terms, what are the benefits for a given brand?

I believe there are many benefits. Let's take the automobile as an example: the different models using the above techniques allow us to set up scorings of intention to buy a new vehicle, detect the churn (and prevent it), predict the mileage of a vehicle, its maintenance, the ideal time of its renewal, set up product recommendation engines and therefore quickly optimize the campaigns as much as possible. This enables a brand to reduce unnecessary expenses on AdWords, Facebook, etc. by targeting consumers more closely (the result being a stronger commercial impact but a reduced marketing budget). We then work on real "life moments" for prospects and customers of the brand to offer them the right product with the right offer, at the right time and on the right channel...


On the other hand, automating the production and distribution of a campaign makes it possible to reduce purely operational tasks: an algorithm can manage 1000 targets in same campaign, which a human being could never do. This provides a real gain in efficiency and time that allows customer marketing teams to focus on strategy.

 

Does this progress also extend to the final consumer?

Algorithms make it possible to better target when to talk to consumer, which products to recommend (recommendation engine), via which channel (email, Facebook, sms, etc.). This transforms mass marketing into a more acceptable personalized relationship which is more meaningful to consumers. A real relationship, much as the one a storekeeper would have had with his customers 50 years ago. In the final analysis, if all is done well the consumer is a winner since while we may even save the consumer some time by offering what is needed, when it is needed.

 

What other applications can we expect in the future?

I think we will see a progression of what we are doing currently, but with greater fine tuning and pertinence: better targeting, improved recommendations, and optimization of increasingly sophisticated campaigns. This can then be applied to other fields. For example, the "creative" part could also be managed by an AI that would propose an email template, image content and the optimal wording per target, or ultra personalized advertising spots with hundreds of automatically generated versions... The future still holds some nice surprises for predictive marketing!

 

 

Topics: Interview d'expert, Predictive Marketing, AI &Data

Back to Home Page