Chapter 1 Introduction
1.1 Background and significance
Email remains a low-cost and dependable communication channel, but the same openness allows unsolicited, deceptive, and malicious messages to reach users. Static keyword rules and blacklists are easy to inspect, yet senders can evade them by changing spelling, inserting irrelevant characters, and shifting vocabulary. Reviews of spam filtering consequently identify content change and concept drift as continuing limitations of fixed decision rules [2-3]. This study treats spam detection as a supervised learning problem and evaluates several classic classifiers under one reproducible protocol.
1.2 Research problem
Machine learning does not remove the need for careful evaluation. False positives may hide legitimate email, whereas false negatives leave spam in the inbox. Accuracy alone cannot describe that trade-off, and a result obtained on one historical corpus cannot be generalized to all languages, periods, or mail systems. The study therefore compares logistic regression, support vector machine, and random forest with accuracy, precision, recall, F1, ROC-AUC, confusion matrices, and inference efficiency [8-10].
1.3 Scope and contribution
The empirical work uses the UCI Spambase dataset, which contains 4,601 observations, 57 numeric features, and a binary spam label [1]. The implemented system accepts this fixed feature representation; it is not presented as a universal parser for raw email. The contribution is a traceable comparison and a working Flask prototype rather than a new classification algorithm.
1.4 Organization
Chapter 2 introduces the theory and evaluation measures. Chapter 3 translates the research task into system requirements and architecture. Chapters 4 and 5 describe model and application implementation. Chapter 6 reports the experiment and system tests. Chapter 7 summarizes the findings and limitations.