Explainable AI

 

A new field of study known as explainable AI seeks to make AI results easier to understand. A problem with algorithms using regression analysis on large datasets is knowing how the results were obtained.  Neural network algorithms adjust weights internally to achieve the highest correlation to determine the results. The issue is understanding which characteristics of the datasets are most significant.

For example, AI software might indicate that a risk is missing from the risk plan, more training is required for two resources, or a specific change request is highly probable. A stakeholder asks why, and the explanation is not readily available. Explainable AI is a field of study that seeks methods to determine how AI works and how to make the results more transparent.

Here are some methods I use to analyze AI results.

1. Manipulate the features. This is similar to performing a simulation but running the algorithm several times with slightly different data to determine what data is most relevant to the results.
2. The number of datasets used in project management to make predictions is usually smaller than in other disciplines. For smaller datasets, I use a genetic algorithm to determine the data features that have a higher impact on the results. Typically, the software shows several features and indicates the factors that account for over 90 percent of the output.
3. The last possibility is to change the project datasets by adding, reducing, or using different datasets. This is similar to my first point but requires a focus on complete datasets instead of data fields.

Why does it matter?

There are several reasons to be concerned about AI delivering results that project managers don’t understand. One of the reasons is ethical issues such as biased data and determining how bias affects the results. As AI becomes ingrained in project decision-making, explainable AI is required to detect and explain abnormal results. Academics and programmers continue to progress with different approaches, such as measuring the level of trust in the result.

Conclusion

As project managers learn to collaborate with AI software, explainable AI will play a critical role in building confidence that decisions can be taken based on AI results.

References

Hoffman, R. R., Mueller, S. T., Klein, G., & Litman, J. (2018). Metrics for Explainable AI: Challenges and Prospects.

Feiyu Xu, Jun Zhu, Wei Fan, Yangzhou Du, Hans Uszkoreit, & Dongyan Zhao. (2019). Explainable AI: A Brief Survey on History, Research Areas, Approaches and Challenges.