E-TİCARET PLATFORMLARINDA AÇIKLANABİLİR YAPAY ZEKAYI KULLANARAK MÜŞTERİ KAYBINI TAHMİN ETME: MÜŞTERİ TUTMA STRATEJİLERİNE PAZARLAMA ANALİTİĞİ YAKLAŞIMI

Bilici F.

ULUSLARARASI YÖNETİM BİLİŞİM SİSTEMLERİ KONGRESİ , Antalya, Türkiye, 15 - 17 Ekim 2026, ss.1-17, (Tam Metin Bildiri)

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Basıldığı Şehir: Antalya
  • Basıldığı Ülke: Türkiye
  • Sayfa Sayıları: ss.1-17
  • Bursa Uludağ Üniversitesi Adresli: Evet

Özet

In e-commerce, customer churn is a strategic marketing problem due to the inability to directly observe customer abandonment in non-contractual relationships and the increasing reliance on predictive analytics for retention decisions. This study develops a current and reproducible explainable artificial intelligence approach to predict customer churn from a marketing and CRM perspective and to explain the reasons for these predictions. Publicly available data generation code for CRROS customer behaviour simulation covering the period January 1, 2024 – March 31, 2026, was reapplied with fixed random seeds, and customer-level modelling data was generated from 8,000 customer profiles, 250,000 interactions, and 24,520 transactions. Following the source modelling logic, customer churn was defined as remaining above the 80% threshold of the time-to-last purchase distribution, and 1,595 customers (19.94%) were classified as churn. To prevent target leakage, recency variables, which directly define churn (customer loss), were removed from the purchase-status model's inputs. Logistic Regression, Random Forest, and XGBoost were compared using stratified five-fold cross-validation and a hold-out test sample separated from the 20% training period, and Accuracy, Precision, Recall, F1, ROC-AUC, and PR-AUC were reported. SHAP generated explanations at both the global and customer levels. In the hold-out results, Random Forest yielded the highest accuracy (0.668), Logistic Regression the highest customer churn sensitivity (0.881), and XGBoost the highest ROC-AUC (0.560). The global SHAP analysis highlighted total engagement, engagement/purchase ratio, monetary value, and purchase frequency. However, the inclusion of non-actionable variables, such as age and location, in the explanations indicates that XAI should be used not only for strategy generation but also for model governance and suitability auditing. The study proposes a Predict–Explain–Verify–Act cycle and suggests that data quality and the definition of customer churn may be more decisive than algorithmic complexity.