Dhara Solanki1, Kush Patel2 1Department of Computer Science, CSPIT, CHARUSAT, Changa, India 2Department of Information Technology, GCET, CVM University, Anand, India
Modern agriculture operates under severe information asymmetry: field-level sensor data, satellite imagery, and soil chemistry readings remain siloed across individual farms, preventing the large-scale learning that machine learning models need to generalise well. At the same time, farmers and agronomists rightly expect to understand why a model recommends a particular action before they act on it. This paper presents FedXAI-Agri, a federated learning architecture augmented with SHAP-based and LIME-based explainability modules, designed to predict crop yield and detect plant disease without transferring raw farm data to any central server. A federation of twelve simulated farm nodes contributes local gradient updates; a central aggregator applies FedAvg with differential-privacy noise clipping. The global model reaches 94.7% classification accuracy on a held-out disease dataset and reduces mean absolute error in yield prediction to 3.12 quintals per hectare, a 21.4% improvement over a centrally trained baseline trained on the same aggregate dataset. SHAP global feature rankings show soil nitrogen, relative humidity, and leaf temperature as the three most influential variables, while LIME local explanations restore full interpretability at per-prediction granularity. The framework runs at inference on a Raspberry Pi 4, making it viable for lowconnectivity edge deployments typical of rural India and sub-Saharan Africa.
Federated Learning, Explainable Artificial Intelligence, Smart Agriculture, SHAP, Differential Privacy, Crop Yield Prediction, Plant Disease Detection
Çağrı Çakır Graduate School of Science and Engineering, Özyeğin University, Istanbul, Türkiye
Debt collections strategies in banking rest on assumptions that are rarely tested empirically—among them, that escalating contact intensity and communication harshness reduces delinquency at acceptable retention cost. This study examines those assumptions using association rule mining (ARM) alongside a conventional logistic-regression scorecard. Using 64,016 collection-action records covering 14,802 delinquency episodes from 8,965 customers, we transform episode histories into transactional “baskets” that capture (i) the policy footprint of intensity–tone–delinquency staging, (ii) responsiveness-adjusted behavioral regimes, and (iii) regimes further enriched with risk, profitability, and delinquent amount. ARM uncovers strongly structured, policy-driven escalation pathways; however, post-mining outcome evaluation shows that customer responsiveness—interacting with escalation rather than treatment severity alone—is a dominant discriminator of churn risk within the same nominal regime. Contrary to the conventional expectation that engagement is protective, within harsh, late-stage escalation regimes it is the responsive customers who churn at elevated rates (22.4% versus 14.4%). Class-conditional analysis further indicates that churn is concentrated in an escalation-despite-engagement configuration, and that this mechanism varies systematically by customer value tier. A benchmark logistic-regression scorecard achieves only moderate discrimination (AUC ≈ 0.63), highlighting the limits of additive models for capturing interaction-driven churn mechanisms. Because actions are policy-driven and key variables are episode-level summaries, the results are interpreted as associative/post hoc patterns rather than causal effects of treatment choices.
Debt Collection, Customer Churn, Association Rule Mining, Customer Retention, Collection Scorecard
Copyright © SIGEM 2026
