International Conference on Computer and Knowledge Engineering

Home / 15th International Conference on Computer and Knowledge Engineering

Adaptive Pattern Reconstruction Using Linear Regression for Improved TPS Anomaly Detection

Authors :

Ali Azarsina¹ Alireza Safarzadeh² MohammadReza Jamali³ Abdolhossein Vahabie⁴

1- University of Tehran 2- University of Tehran 3- pulseware co 4- University of Tehran

Keywords :

Anomaly Detection،Non-Homogeneous Poisson Process،Trend‑Aware Regression

Abstract :

The increasing reliance on electronic banking has made service availability and fault detection critical concerns for financial institutions. Even brief service disruptions can lead to failed transactions, customer dissatisfaction, and significant financial loss. Accurate and timely detection of service anomalies is thus vital for operational resilience and customer trust. We present a method to quantify lost transactions and compute perceived availability in banking switches by reconstructing expected transaction-per-second (TPS) demand patterns. Outage and load-drop intervals are first detected by comparing the target day’s TPS wave against a reference wave built from neighboring days. We then model both real and expected TPS using a Poisson framework to generate demand estimate patterns. The gap between the reconstructed and observed waves yields counts of missed transactions and an end-user–centric availability metric. Evaluation on real bank data demonstrates that our approach effectively uncovers faults and provides accurate availability estimates. Compared to a baseline method, our technique reduced the mean squared error by 9.1\%, and lowered false positive rates in anomaly detection from 13.3\% to 8.7\%, while maintaining high sensitivity.