This year's edition of AdNovum's breakfast event IT@8 was dedicated to Machine Learning. Dr. Kristjan Vassil of xUpery and Dr. Tom Sprenger of AdNovum shared some theoretical and practical insights.
On Thursday, June 15, AdNovum welcomed numerous representatives of the finance and insurance industry to its annual breakfast event IT@8. They gathered at the Clouds, the restaurant with a panoramic view 120 meters above Zurich, to learn about the latest trends and best practices in self-learning software.
Theory and practical experience
Following an introduction by AdNovum's CCO Peter Gassmann, guest speaker Dr. Kristjan Vassil of xUpery shared some theoretical and practical insights into the areas of Artificial Intelligence and Machine Learning.
To start with, he explained the meaning of machine learning, what it requires and when its use makes sense. One prerequisite is the availability of data and to know what they should be used for. They could, for example, be used in connection with customers who are interested in a specific banking product.
Machine learning consists of supervised learning that replicates regularities, and unsupervised learning where target values are unknown in the beginning and that tries to identify patterns in the data entered. Using the example of a bank, Dr. Vassil illustrated how supervised learning enables sales predictions. By means of a case of anomaly detection, he explained unsupervised learning.
Lessons for users of machine learning
In his speech, AdNovum's CTO Dr. Tom Sprenger made it clear that the tool per se is not the solution. However, a suitable machine learning toolbox is key. He presented 6 lessons that are crucial to use machine learning successfully:
Lesson 1: the Machine Learning Project Cycle
In addition to an in-depth understanding of the area, it requires data access, screening and pre-selection. Also part of the project cycle are data interpretation as well as consolidation and implementation of the results. It is important to understand that machine learning projects require an iterative approach in order to achieve suitable results.
Lesson 2: Data Access and Training Data Quality
One challenge machine learning faces is the fact that there is no universal data access. Even a knock-out criterion are incomplete, inconsistent or ambiguous data.
Lesson 3: Domain Knowledge
Experts wanted! For feature engineering, optimizing machine learning algorithms and validating and interpreting the results.
Lesson 4: the Appropriate Machine Learning Approach
To identify it, the following questions need to be answered: analytics vs. real-time analytics? Supervised vs. unsupervised anomaly detection?
Lesson 5: Be Aware of the Impact
Think of the impact already when designing the machine learning solution. While, for example, wrong recommendations by companies such as Netflix or Amazon have only a minor impact, false positives and, in particular, false negatives may have severe consequences. For example when access to a protected resource, such as a bank account, is granted to an unauthorized person.
Lesson 6: Start Small and Grow
Meaning: Set realistic goals and consider factors, such as data selection, complexity of the model, tool selection and system size.
When is a machine learning solution ready to use?
The starting point for machine learning are data. For them to be usable in the self-learning process, they need to meet a certain quality standard. In the process, different static algorithms are used to answer the following questions:
- Is there a specific question I need to answer based on a variety of data?
- What kind of information do the data provide and what added value can they generate for my business?
Data are normally separated: one part is used to train algorithms, the other part to validate the results. Once the selected machine learning model delivers decent results, it can be brought live and used to handle real-time data such as predictions and anomaly detection. Artificial Intelligence even allows selecting the most suitable algorithm in realtime. Now the machine learning solution is ready to be integrated into the existing application landscape.
Provided the model is optimized on a regular basis, i.e., every 3–6 months, machine learning can unlock its full potential.
Event series IT@8
The breakfast event dedicated to Self-Learning Software was part of AdNovum IT Consulting's series "IT@8". The series is aimed at managers and addresses current trends and their impact on IT and the business. We usually choose one trend and discuss its impact with selected guest speakers who have in-depth practical experience.
The event takes place once a year; the next one is planned for spring 2018. Please send us an e-mail to firstname.lastname@example.org if you would like to receive an invitation.