DISA

• When? November 4^th 14-16
• Where? D1140 – Växjö (link will provided for those who wants to attend online)
• Registration: We would like to know how many that will attend onsite/online in order to get some fika for those onsite. So please register by November 2^nd https://forms.gle/ZwwgoQ4JK4e41BBR6

 Agenda

14.00-14.10     Welcome and practical information from Welf Löwe

14.10-14.55     Presentation and discussion: Visual Analytics for Explainable Machine Learning in a Nutshell – Angelos Chatzimparmpas

14.55 – 15.05  Coffee break

15.05 – 15.50  Presentation and discussion: Getting the most out of health data, combing the best of two worlds – Olle Björneld

15.50 -16.00    Sum up and plan for our next seminar on January 13^th

Abstracts

Visual Analytics for Explainable Machine Learning in a Nutshell – Angelos Chatzimparmpas

Machine learning (ML) research has recently gained much attention, with various models proposed to understand and predict patterns and trends in data originating from various domains. Unfortunately, users find it harder to evaluate and trust the results of these models as they become more complex because most of their internal workings are kept in secret black boxes.

One possible solution to this problem is the explanation of ML models with visual analytics (VA) since it enables human experts to analyze large and complex information spaces such as data and model spaces. By doing so, evidence has shown an improvement in predictions and an increase in the reliability of the results.

This talk aims to provide an overview of the state-of-the-art in explainable and trustworthy ML with the use of visualizations, as well as the development of VA systems for each stage of a typical ML pipeline. Furthermore, we will briefly introduce some of these tools and discuss how such VA techniques can help us not only understand ML models but also do this in a human-centered and steerable way.

Getting the most out of health data, combing the best of two worlds – Olle Björneld

Machine learning driven knowledge discovery on real world data based on domain knowledge. Real world data does not comply with machine learning models very well and prediction models perform suboptimal if pre-processing of data is deficient.

Based on experience from medical registry studies using electronic health data (EHR) performed in collaboration with domain experts, data analyst and statistician an automatic feature engineering framework and method have been developed. The framework is called automatic Knowledge Driven Feature Engineering (aKDFE) and have been evaluated by machine learning pipeline.

Experiment shows that prediction models performs better if aKDFE is used without losing explainability, but more experiments need to be performed in other domains to fully quantify the results. The key aspect is how to concentrate and mine inherent knowledge in transaction data to optimal machine learning driven prediction models.

A warm welcome,

Welf & Diana