Importance of data (and humans) for machine learning

I was last week in London for the Marketing Summit organized by Adobe. A 2017 edition resolutely turned towards machine learning. If no one can dispute the importance of machine learning in the service trades (the ” why? “), As with every new major innovation, the greatest confusion reigns over the exact perimeter (the ” what ? “) And implementation (the” how? “). Whether Adobe, SalesForce or IBM, all major publishers, and even small ones, we use the machine learning in their speech by explaining that their artificial intelligence is smarter than that of others. The problem is that there is no good or bad AI, it all depends on the quality of the datasets and the interpretation of the results by the teams.

A forced marriage between marketing and AI

Artificial intelligence is definitely the topic that everyone is talking about right now, and that is the theme of all events, from SXSW to Marketing Summit last week. The reason for this craze is simple: we, marketing professionals, do not want but need artificial intelligence. It has not escaped you that the proliferation of media poses serious problems for communication professionals and that the multiplication of channels also poses great problems for marketers: as targets and customers disperse, resources and power advertisers are diluting.

I had already discussed the subject at the beginning of the year ( Is the AI the future of marketing? Certainly … ) and remains convinced that AI and their capacity for creative destruction are at work: they accelerate the disappearance of low added value trades and at the same time open up many opportunities. You can see the glass half empty ( artificial intelligence: the end Gong for media agencies? ) Or the half-full glass ( Artificial Intelligence Market Forecasts ), but in all cases, the share of work done by artificial intelligence robots in our everyday work will only increase ( Nearly 80 percent of US display ad spend will be programmatic in 2017 ).

The best way to approach this fact is to say that AI is precisely the solution to the fragmentation of the audience and the multiplication of media: what we lost in power (the TV), we must compensate it in precision. And it is precisely here that the AI and more particularly the tools doped with the machine learning will be particularly useful.

So no, this is not the end of media agencies, but rather the end of an era for media agencies that abandon manual work (coordination and confirmations made by phone or email) to focus on automation. But do you say that in all cases, it will take humans to set up, monitor, analyze, arbitrate …

In “artificial intelligence”, there is “artificial”

As stated in the introduction to the article, there is a great deal of confusion about what artificial intelligence is, what they can do, and how best to exploit them. To summarize a long explanation, artificial intelligence is a computer tool that simulates human intelligence. This definition applies to both vocal discs used for decades and chatbots, it is the same principle.

Would you say that a hammer is stupid or intelligent? Neither, he is as intelligent (or stupid) as he who handles it. It’s the same for an artificial intelligence: the important thing is not the choice of AI, but the way you will exploit it. In an integrated environment like Adobe or SalesForce cloud marketing, it works well, but as soon as you start tinkering and stacking technical layers, it gets complicated.

I draw your attention to the fact that there are no ready-made solutions, except in certain very specific areas such as natural language processing (NPL) or image recognition. Exploiting an artificial intelligence necessarily involves a long learning phase to collect the right data, recruit the right skills, exploit the right algorithms …

A company will have to go through different stages of artificial intelligence :

  • Assisted intelligence that improves the productivity of existing tasks (eg real-time bidding);
  • augmented intelligence that enables tasks that are not only possible with human staff (eg personalization of individual messages);
  • autonomous intelligence that will no longer require human operators (eg Google smart display campaigns )

Again, the problem does not come from AI itself (there is no good or bad AI), but shortcuts that abuse abusers sellers of miracles. Artificial intelligence, and more so machine learning, are complex disciplines that require a minimum of pedagogy and especially a lot of preparation. In summary, this requires a strategy (see A Strategist’s Guide to Artificial Intelligence ).

In “machine learning”, there is “learning”

Machine learning is one of the branches of artificial intelligence. It is a discipline that appeared in the 80s with the first research work on statistics and neural networks. To make it simple: machine learning is used to develop learning processes that allow a machine to evolve, without its algorithms being modified, with the aim of building a predictive model. Depending on the context and the data available, a machine can learn in several ways (supervised, unsupervised, reinforced, transfer …) and according to several approaches (decision trees, linear regression, neural networks, Bayesian networks, vector machines …).

The application domains of machine learning are numerous but revolve around data analysis and their restitution in the form of classifications (with already known classes), clusters (with empirical clusters), predictions, density estimates … Adapted to the marketing professions, machine learning makes it possible to better target customers (micro-segments), to improve the impact (real-time analysis and automatic arbitrage), to customize offers and messages, to write personalized content, recommend products, calculate more precisely the contribution of a particular medium during a campaign …

In short, machine learning is a genuine performance accelerator based on the analysis of large datasets. That’s good, because the production of data has grown exponentially in recent years, and because the technological bricks to make big data have become much more affordable and better mastered. In synthesis: all planets are aligned to explore the uses of machine learning.

The problem is that machine learning algorithms are not directly applicable to marketing, communication or online sales. Unsupervised learning techniques (deep learning) are perfect for analyzing very large amounts of heterogeneous data and extracting the strongest correlations, but not necessarily to tell you what to do. These algorithms are used to identify patterns, weak signals from correlations, but if there is not enough data or if it is not the right data, then the correlations will be wrong. This is where I think the market is wrong: we are promised mountains and wonders by boasting the merits of this or that AI, while they use almost the same algorithms (from academia). The difference will rather be in the richness of datasets (“data sets”), the prioritization of one algorithm over another (since we can combine the methods of analysis) and how the teams will interpret the data sets. results.

To get a good understanding of the work necessary to set up an AI, I offer you this very good feedback from Crédit Mutuel on the implementation of IBM’s AI to assist customer service: “No, I ‘IA Watson is not magic’ We learn that the learning phase still lasted 10 months and monopolized between 10 and 15 people (not full-time, but a great team all the same). The project managers said that it was necessary to involve business experts to swallow 6,000 standard questions and deduce 4,000 rules and decision trees to 2.00 branches that must be regularly updated (groups!). But obviously, the game is worth the candle, because the goal is to speed up the processing of incoming questions: ” If we save 5 000 days a day to 20,000 account managers on the processing of emails, we will quickly make money from ‘investment ‘. Indeed, there is a cost, which will be quickly reimbursed by the performance gain, but the costs and implementation times are not negligible.

We touch here the heart of the debate: there is a limit in the work of machine analysis, in the end, the interpretation of the results should be done on the basis of experience or intuition. You may have the greatest computing power (at IBM, Amazon or Google), if you make a mistake in your interpretation, then all this work will be reduced to nothing. Men and machines must work together to provide the best results. Read on this topic this very good article that compares AI to micro-services: Artificial intelligence, brilliant and stupid.

Morality: your priority is not to choose the best algorithm, but to gather enough data and surround yourself with the right people, both supervised and semi-supervised learning specialists, but also experts in the field that you want to improve. We have here the demonstration that the AI will actually destroy jobs with low added value, but create others, with high added value, and especially that the issue is at the data level. From this perspective, Google’s acquisitions of many years ago are better understood (see From King Content to Queen Data released in 2010), and the ongoing battle around health data ( The Terrifying Black Market for your Private Medical Records ).