Postdoctoral Position in Counterfactual Explanations for Multivariate Time Series

Counterfactual explanations answer the question "what would have to change in the input for the model to predict something else?" [1]. They are intuitive and point to possible actions. They have been studied extensively for tabular data, where methods balance proximity to the original instance, sparsity, diversity and plausibility [2]. For time series, research has mostly focused on classification [3, 4, 5]. A recent large-scale evaluation of sixteen such methods [6] found that none achieves validity, proximity, sparsity and plausibility at the same time, and called for domain specific constraints to make counterfactuals more realistic. Counterfactual explanations have also been proposed for forecasting [7, 8, 9], anomaly detection [10, 11, 12, 13] and survival analysis [14, 15] tasks. Several issues remain open across these tasks. The first is plausibility, given that the generated counterfactuals often break relations that hold in real data. Data management offers a principled tool for this in the form of integrity constraints. Denial constraints were recently used to project counterfactuals for tabular data onto realistic regions [16], and temporal constraints such as speed constraints are already used to clean time-series streams [17]. The second issue concerns the control users may have over the counterfactual generation process. In practice, only some of the input time series can be acted on, during certain periods and within certain ranges. Existing methods tackle this issue only partially or in task-specific ways. For example, [18] learns a small number of time intervals in which a classification counterfactual may differ from the original. Moreover, [7, 8, 9] bound the forecast or the actionable inputs, while [13] constrains how anomalies propagate over time. Recent work argues that counterfactuals for time series should be understood as temporally coherent interventions [19]. A third issue is related to the granularity of explanations. Most methods propose counterfactuals for one instance at a time. Global or group counterfactuals, studied for tabular data [20] and recently for time-series clustering [21], can instead show how an entire population of series would need to change. Survival analysis tasks raise additional questions. Existing counterfactuals for survival models either use static covariates [14] or change the observed window of a multivariate series [15]. Both define the target on a single number that summarises the predicted survival curve, such as the mean time-to-event or the remaining useful life. However, this aggregation effectively turns the problem into one of counterfactual regression rather than counterfactual survival analysis. Moreover, time-to-event predictions may also depend on future covariates (i.e., feature values after the current/historical state). Finally, causal machine learning techniques [22, 23] aim to estimate outcomes under hypothetical future treatment sequences. In this context, the use of future covariate paths as the object of a counterfactual explanation has received little attention, partly because few models take future covariates into account.

Subject The aim of the postdoc position is to investigate novel counterfactual generation techniques for multivariate time-series models that are plausible with respect to the input data, controllable by the user and applicable beyond classification tasks. A guiding idea is to describe a counterfactual through two key elements: a target on the model output, and a set of constraints on the input. The constraints specify which variables may be modified, during which periods, within which value ranges, and which relations must be preserved. The same principles can then serve different tasks, with the task entering mainly through how the target is defined. Research directions include, but are not limited to, developing techniques that incorporate temporal integrity constraints, such as denial constraints, into counterfactual generation and provide users with explicit control over where, when, and how much a series may change. For survival models, the explanation should target changes to the predicted curve rather than to a single distribution summary, by considering counterfactuals over future covariate paths. Group counterfactuals, which apply one common change to a set of series such as a fleet of machines or a cohort of patients, are another possible line. The outcomes of this research are intended to support decision-making, either in human-in-the-loop settings or through integration into agentic pipelines.