THANK YOU FOR SUBSCRIBING
Bringing Machine Learning to your Business
By Dr. Sébastien Foucaud, Vice-President, Data Science, XING
In addition, with a few exceptions (e.g. autonomous driving), the current business solutions presented by ML are limited to the improvement of features (e.g. increased efficiency of marketing campaign) or enabling new features (e.g. search image by content rather than keyword) of existing products.
In this article I will describe elements necessary to widen and deepen the use of ML based on my experience at XING.
1. Product in Business Context: Understand the Problem to Solve.
Don’t get it wrong, the difficulty with ML is not technology, but how to apply it in a relevant way to the business case: requires understanding of the problem to be solved and identifying the business objectives. Defining the correct KPIs and SLAs for each ML project is essential to ensure value to the business and proper optimization (e.g. Click-Through-rates and Active users).
However, as currently the most common way of using ML is to improve existing features, benefits and user behavior are usually already known. It is therefore “just” a question of knowing which ML solution to use for which problem.
2. Data Acquisition and Preparation are Key: Which Data is Needed for What.
Once the business objective and desired performance are set, the next immediate step is to evaluate if the data required for training your model already exist (labeled data in context of supervised ML), is of relevant quality, needs to be enriched (e.g. via active learning techniques) or needs to be acquired fully.
One critical aspect to bear in mind is that in real-world datasets some events (usually your event of interest) are extremely rare and requires special care (class imbalance).
Integrating ML based solution into your product environment and applying best software development practice (unit test, version controlling, etc.) is essential
Integrating ML based solution into your product environment and applying best software development practice (unit test, version controlling, etc.) is essential.
It also requires a specific twist as even with the same code and infrastructure, the corpus of data on which the solution is applied will be different, hence your model need to be retrained. In addition to versioning the code you may want to version the training and test data sets used. In general, heavy usage of containers is required (e.g. Docker).
4. Get User Feedbacks at all Costs!
Implementing a feedback loop is critical, especially given that almost all solutions are supervised ML based. As in real world user behavior and (product) environment changes, your data will change with time.
Implicit feedback (e.g. clicks) is easier to collect, but explicit feedback (e.g. rating) brings more value. It is also important to integrate the feedback query into the user journey and experience to ensure its relevance.
5. No model Fits all your Cases: Think Explainable, Simple and “Ensemble”.
Although performance and accuracy of the model is important, in the real world reliability and robustness of your system, along with transparency and explainability matters more! You should be able to provide to your user explanation of how a decision is arrived at (e.g. rejection of a mortgage) and your system needs to work for all users, all the time.
This is why the newest, state-of-the-art approaches are rarely the best choice. A collection of more simple algorithms, merged together following ensemble techniques, will help you in managing the duality of accuracy/explainability.
6. Automating of ML Production
As described above ML requires constant tuning of ensemble of models with many parameters based on your ever-changing data, and needs to be reliable and robust enough for production. All these stages can be automated: the selection and build-up of your training and test data sets, feature engineering, and the hyperparameters (Bayesian) optimization of your ensemble of models. This is an essential step for transparency (full system versioning) as every step can be systematically reproduced.
7. The Rise of AI Boutiques
As a large part of the processes described above are open-sourced and with access to efficient cloud-based infrastructure and large computation power, technology is rarely the showstopper for building ML solutions, but rather understanding the business case and the data available, along with the know-how of implementing solutions at scale. This is spurring the growth of small, but highly differentiated service providers who can support companies to accelerate their journey to ML by providing very niche and specific know-how in business and technology, but with a much lower cost base than conventional large service providers.
Sharam Hekmat, CIO, Ioof Holdings, Australia
Why Artificial Intelligence Will Empower the CIO?
Gregory B. Morrison, SVP & CIO, Cox Enterprises
How to Get to AI-first
Ani Paul, CIO, ING Australia
AI and the Future of Field Service: Moving from Efficiency to Innovation
Michael Alcock, Director-CIO Executive Programs & Content, Microsoft [NASDAQ: MSFT]