Seminar I The Performance-Interpretability Tradeoff? Not a problem for discourse analysis.

The performance-interpretability trade-off (PIT) is a key consideration in data analytics. It is the trade-off between how well a model performs and how interpretable it is to humans. Across applications in healthcare, business, and so on, more complex models tend to perform better in decision-making tasks but are harder to interpret, and vice-versa. PIT is often seen as a problem to be addressed by improving algorithms or finding a pragmatic ‘balance’ between performance and interpretability.

In this talk, I share initial thoughts on a recent project about PIT and discourse analysis – a field that for various reasons is not often associated with data analytics. My main argument is that PIT is not a problem in discourse analysis, but a window of opportunity. At the most general level, the inherent multivalence of discourse analysis would already reject any P-I dichotomy for a more nuanced view of what it means for a model to ‘perform’ and ‘interpret’. And for actual discourse data, P and I have different ‘elasticities’ and are not always negatively correlated. Quantitative changes in P are not straightforwardly linked to changes in I. Discourse analysis is thus an excellent example of how domain knowledge must be allowed to shape data analytics. To support these highly speculative ideas, I provide some preliminary results that compare P vs. I among competing time series models of social media discourse across time.