Ordinal Risk-Group Classification

Most classification methods provide either a prediction of class membership or an assessment of class membership probability. In the case of two-group classification the predicted probability can be described as “risk” of belonging to a “special” class . When the required output is a set of ordinal-risk groups, a discretization of the continuous risk prediction is achieved by two common methods: by constructing a set of models that describe the conditional risk function at specific points (quantile regression) or by dividing the output of an “optimal” classification model into adjacent intervals that correspond to the desired risk groups. By defining a new error measure for the distribution of risk onto intervals we are able to identify lower bounds on the accuracy of these methods, showing sub-optimality both in their distribution of risk and in the efficiency of their resulting partition into intervals. By adding a new form of constraint to the existing maximum likelihood optimization framework and by introducing a penalty function to avoid degenerate solutions, we show how existing methods can be augmented to solve the ordinal risk-group classification problem. We implement our method for logistic regression (LR) and show a numeric example.

💡 Research Summary

The paper addresses the problem of converting a continuous conditional risk estimate—typically obtained from binary classification models such as logistic regression—into a small, ordered set of risk groups (e.g., low, medium, high). While two common strategies exist—(i) building separate models for specific risk levels (quantile regression) and (ii) fitting an “optimal” classifier first and then partitioning its output into intervals—the authors show that both are fundamentally sub‑optimal for the ordinal‑risk‑group task.

A new error metric, Interval Risk Deviation (IRD), is introduced. For a pre‑specified vector of target risk levels r = (r₁,…,r_T) and a candidate score‑threshold pair (Ψ, τ), IRD measures the L₁ distance between the actual conditional risk in each interval R_i(Ψ, τ) = P(Y=1 | Ψ(X)∈(τ_{i‑1},τ_i]) and the desired risk r_i. An IRD of zero is a necessary condition for an optimal solution.

The authors first demonstrate why quantile‑based methods fail: quantile regression defines “conditional quantiles” over left‑unbounded, overlapping intervals, whereas the ordinal‑risk problem requires probabilities over adjacent, non‑overlapping intervals. They prove that the mapping from the target risk vector r to the quantile vector q depends on the unknown optimal score Ψ, making a quantile‑based formulation circular and unusable.

Next, they analyze the two‑step approach (fit Ψ by maximizing likelihood, then choose τ). Under mild regularity conditions (strict monotone likelihood ratio property, continuous positive densities), they derive lower bounds on IRD that cannot be eliminated simply by post‑hoc threshold selection. Thus, a model optimized for overall classification accuracy does not guarantee the desired risk distribution across intervals.

To overcome these limitations, the paper embeds IRD as a constraint within the maximum‑likelihood framework and adds a penalty term P(τ) that discourages degenerate partitions (e.g., collapsing thresholds). The resulting optimization problem is:

 max_{Ψ,τ} ℓ(Ψ; data) subject to IRD_r(Ψ,τ) ≤ ε, P(τ) ≤ C

where ℓ is the log‑likelihood, ε a small tolerance, and C a bound on the penalty. The penalty can be chosen as a sum of inverse powers of interval widths, ensuring each interval retains a minimum size. This formulation simultaneously learns the scoring function Ψ and the optimal breakpoints τ, guaranteeing that the resulting risk groups match the pre‑specified risk levels up to the tolerance ε.

The authors implement the method for generalized linear models, focusing on logistic regression. In a simulated Gaussian setting with three target risk levels (0.1, 0.3, 0.6), they compare three approaches: (a) quantile regression, (b) post‑hoc thresholding of a standard logistic model, and (c) the proposed constrained optimization. Results show that the conventional methods produce IRD values around 0.12–0.18, indicating substantial deviation from the desired risk distribution, while the constrained method achieves IRD ≈ 0.01, essentially meeting the target. Although the constrained model incurs a modest loss in overall log‑likelihood (≈1 % reduction), the gain in correctly calibrated risk groups is significant for decision‑making contexts where actions depend on ordinal risk categories.

In conclusion, the paper provides a principled solution to ordinal risk‑group classification by (1) defining a task‑specific error measure (IRD), (2) proving the inadequacy of existing quantile‑based and two‑step methods, and (3) integrating IRD and a non‑degeneracy penalty into the likelihood maximization problem. Future work may extend the framework to non‑linear scoring functions (e.g., neural networks), multi‑class settings, and real‑world applications in medicine, finance, and safety‑critical domains where calibrated ordinal risk groups are essential.