An Explainable Chatbot for Student Orientation in Morocco: SHAP and LIME on a Hybrid Neural Network-SVM Model

Abstract

Poor orientation feeds school dropout in Morocco. Leaving the scientific common core (tronc commun), a pupil must pick one of three tracks: Mathematical Sciences (SM), Physical Sciences (SF), or Life and Earth Sciences (SVT). Software can help with that choice. Accuracy alone cannot: a model that will not say why it recommends a track earns no trust from the families and counsellors who must act on it. Our system has three parts. A chatbot gathers the pupil’s marks. A sequential hybrid neural network into SVM issues the prediction. An explainability layer built on SHAP and LIME accounts for it, and that layer is the contribution. We trained on some 20,000 anonymised records from the MASSAR system, obtained through the Nador provincial directorate and covering 2018 to 2022. The hybrid reaches 98.87% on the held-out set, against 95.79% for the network alone and 97.52% for the SVM alone, with macro AUC 0.9981 and Cohen’s kappa and MCC both 0.983. The national platform holds over seven million pupil records and cannot be extracted at that scale, so we repeated everything on OULAD, public and far larger, after mapping its variables onto our feature schema. Accuracy is 98.73%. SHAP puts the decision on the three science subjects; LIME accounts for pupils one at a time.