AI-Driven Decision-Oriented Scholarly Influence Forecasting With Calibrated Confidence and Abstention

Forecasting the future influence of scholarly articles can enable earlier, more actionable decisions in research management, editorial triage, and funding oversight. However, most citation forecasting pipelines are evaluated purely as predictive models and often rely heavily on “prestige” proxies (e.g., author history), limiting their suitability as decision-support systems and raising concerns about robustness, fairness, and temporal stability. In this work, we extend our prior AI-CITE framework into a decision-oriented forecasting pipeline that adds calibrated confidence, structured abstention, and explicit controls on prestige dependence—without requiring any additional data collection (AI-CITE2). Building on publication-time inputs, the system fuses neural text representations (title+abstract) with auditable, time-truncated metadata descriptors, and supports both (a) top-quantile influence classification at horizon $H$ and (b) regression of stabilized citation counts via $\log (1+\text {citations})$ . To make forecasts operationally reliable, we introduce calibrated confidence using post-hoc temperature scaling and an abstention mechanism that allows the model to output “I don’t know” when uncertainty is high. We evaluate selective prediction through coverage–risk and selective performance curves, demonstrating how abstention can reduce decision risk while retaining strong ranking ability on the accepted subset. To audit and reduce measured dependence on prestige proxies, we conduct structured prestige ablations and report influence forecasting performance under progressively restricted descriptor sets, alongside metrics that quantify dependence on author-derived features. Finally, we add decision-relevant evaluation beyond standard AUROC/AUPRC, including ranking metrics and temporal robustness checks across publication years, with drift diagnostics that identify distribution shifts affecting confidence and operating thresholds. Across controlled temporal splits with strict identity-leakage prevention, the proposed decision-oriented AI-CITE2 variant maintains competitive forecasting performance while providing calibrated, abstention-aware outputs and reduced measured dependence on prestige-related proxies (as quantified by controlled ablations). These results position citation forecasting as a practical decision-support tool rather than a purely retrospective prediction benchmark.