Evaluating Decision Makers over Selectively Labelled Data: A Causal Modelling Approach

Abstract

We present a Bayesian approach to evaluate AI decision systems using data from past decisions. Our approach addresses two challenges that are typically encountered in such settings and prevent a direct evaluation. First, the data may not have included all factors that affected past decisions. And second, past decisions may have led to unobserved outcomes. This is the case, for example, when a bank decides whether a customer should be granted a loan, and the outcome of interest is whether the customer will repay the loan. In this case, the data includes the outcome (if loan was repaid or not) only for customers who were granted the loan, but not for those who were not. To address these challenges, we formalize the decision making process with a causal model, considering also unobserved features. Based on this model, we compute counterfactuals to impute missing outcomes, which in turn allows us to produce accurate evaluations. As we demonstrate over real and synthetic data, our approach estimates the quality of decisions more accurately and robustly compared to previous methods.

Keywords

Appendices

Appendix 1 Counterfactual Inference

Here we derive Eq. 4, via Pearl’s counterfactual inference protocol involving three steps: abduction, action, and inference [17]. Our model can be represented with the following structural equations over the graph structure in Fig. 2:

$$\begin{aligned} \mathsf {J}&:= \epsilon _{\mathsf {J}}, \quad \mathsf {Z}:= \epsilon _\mathsf {Z}, \quad \mathsf {X}:= \epsilon _\mathsf {X}, \quad \mathsf {T}:= g(\mathsf {H},\mathsf {X},\mathsf {Z},\epsilon _{\mathsf {T}}), \quad \mathsf {Y}:= f(\mathsf {T},\mathsf {X},\mathsf {Z},\epsilon _\mathsf {Y}). \end{aligned}$$

For any cases where $\mathsf {T} =0$ in the data, we calculate the counterfactual value of $\mathsf {Y}$ if we had $\mathsf {T} =1$. We assume here that all these parameters, functions and distributions are known. In the abduction step we determine $\mathbf {P}(\epsilon _\mathsf {H}, \epsilon _\mathsf {Z}, \epsilon _\mathsf {X}, \epsilon _{\mathsf {T}},\epsilon _\mathsf {Y} |j,x,\mathsf {T} =0)$, the distribution of the stochastic disturbance terms updated to take into account the observed evidence on the decision maker, observed features and the decision (given the decision $\mathsf {T} =0$ disturbances are independent of $\mathsf {Y}$). We directly know $\epsilon _\mathsf {X} =x$ and $\epsilon _{_\mathsf {J}}=j$. Due to the special form of f the observed evidence is independent of $\epsilon _\mathsf {Y}$ when $\mathsf {T} = 0$. We only need to determine $\mathbf {P}(\epsilon _\mathsf {Z},\epsilon _{\mathsf {T}}|h ,x,\mathsf {T} =0)$. Next, the action step involves intervening on $\mathsf {T}$ and setting $\mathsf {T} =1$ by intervention. Finally in the prediction step we estimate $\mathsf {Y}$:

where we used $\epsilon _\mathsf {Z} =z$ and integrated out $\epsilon _\mathsf {T}$ and $\epsilon _\mathsf {Y}$. This gives us the counterfactual expectation of Y for a single subject.

Appendix 2 On the Priors of the Bayesian Model

The priors for $\gamma _\mathsf {X},~\beta _\mathsf {X},~\gamma _\mathsf {Z}$ and $\beta _\mathsf {Z}$ were defined using the gamma-mixture representation of Student’s t-distribution with $\nu =6$ degrees of freedom. The gamma-mixture is obtained by first sampling a precision parameter from $\varGamma$() and then drawing the coefficient from zero-mean Gaussian with that precision. This procedure was applied to the scale parameters $\eta _\mathsf {Z},~\eta _{\beta _\mathsf {X}}$ and $\eta _{\gamma _\mathsf {X}}$ as shown below. For vector-valued $\mathsf {X}$, the components of $\gamma _\mathsf {X}$ ($\beta _\mathsf {X}$) were sampled independently with a joint precision parameter $\eta _{\gamma _\mathsf {X}}$ ($\beta _{\gamma _\mathsf {X}}$). The coefficients for the unobserved confounder $\mathsf {Z}$ were bounded to the positive values to ensure identifiability.

$$\begin{aligned} \eta _\mathsf {Z}, \eta _{\beta _\mathsf {X}}, \eta _{\gamma _\mathsf {X}} \sim \varGamma (3, 3), \; \gamma _\mathsf {Z},\beta _\mathsf {Z} \sim N_+(0, \eta _\mathsf {Z} ^{-1}),\; \gamma _\mathsf {X} \sim N(0, \eta _{\gamma _\mathsf {X}}^{-1}),\; \beta _\mathsf {X} \sim N(0, \eta _{\beta _\mathsf {X}}^{-1}) \end{aligned}$$

The intercepts for the decision makers in the data and outcome $\mathsf {Y}$ had hierarchical Gaussian priors with variances $\sigma _\mathsf {T} ^2$ and $\sigma _\mathsf {Y} ^2$. The decision makers had a joint variance parameter $\sigma _\mathsf {T} ^2$.

$$\begin{aligned} \sigma _\mathsf {T} ^2,~\sigma _\mathsf {Y} ^2 \sim N_+(0, \tau ^2),\quad \alpha _j \sim N(0, \sigma _\mathsf {T} ^2),\quad \alpha _\mathsf {Y} \sim N(0, \sigma _\mathsf {Y} ^2) \end{aligned}$$

The parameters $\sigma _\mathsf {T} ^2$ and $\sigma _\mathsf {Y} ^2$ were drawn independently from Gaussian distributions with mean 0 and variance $\tau ^2=1$, and restricted to the positive real axis.