Examining black-box ML models with SHAP

The working group Data Science in the SAA (Swiss Association of Actuaries) is pleased to present its new Meet Up Series: Actuarial Data Science Reading Club. In a casual and interactive setting, actuaries and data scientists will have the opportunity to share their experiences and learn from each other.
Each session will evolve around one practical topic of interest, which will be briefly introduced at the beginning.
If you are looking to discuss ideas, ask questions or simply be inspired by people working on similar topics, this might be for you! The events are free and open to all levels of experience.

Our second meeting will take place on 28.10.2024 (18:00-20:00), at
Edge5 close to the Zürich HB.

Description to enter the room:

```
To access the meeting room, ring the bell at Edge5 / Device Tools or walk in directly *** Door code: 3037# *** Meeting room at Edge5 on the 3rd floor.
```

Examining black-box ML models with SHAP
Marco gives a gentle introduction into SHapley Additive exPlanation (SHAP) and how this method can be used to make black-box machine learning models interpretable.
We show variable attribution to the prediction on the local level of a single observation and talk about different high-level variations of the SHAP algorithm. Specifically, we want to discuss the difference between marginal and conditional SHAP algorithms and their pros and cons. Aggregating local SHAP values one gets a global model interpretation. This shows how single input variables influence the overall prediction. We might also digress into SHAP interaction values and how to interpret the collection of multiple features.
Last, we want to share our own modelling experience and certain shortcomings of SHAP values.

Marco Breitig is Head of Motor Retail Pricing at Allianz Suisse with several years of experience in actuarial modeling and machine learning. He holds a PhD in high-dimensional statistics and abstract probability theory.

If you want to prepare in advance, we recommend to have a look at the following papers:

• Algorithms to estimate Shapley value feature attributions (https://arxiv.org/abs/2207.07605)
• SHAP for Actuaries: Explain any Model (https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4389797)
• Conditional expectation network for SHAP (https://arxiv.org/abs/2307.10654)

Space is limited, so please only register if you are sure, that you will come and unjoin, if you have to cancel!