Smart E-Commerce System with AI Recommendation Engine and Customer Behavior Analytics

E-commerce platforms usually contain a large number of products, which can make it difficult for customers to quickly find items that match their interests. This project presents a Smart E-Commerce System that combines a machine learning recommendation engine with customer behavior analytics and supports the main activities of customers, vendors, and administrators. The recommendation system uses a hybrid approach that combines collaborative filtering and content-based filtering. Collaborative filtering identifies similar product preferences based on user interactions using cosine similarity, while content-based filtering uses TF-IDF to find similarities between product information. When both methods are available, the system combines them using 75% collaborative filtering and 25% content-based filtering. Customer activities are also considered when generating recommendations. Different actions are given different weights based on their importance: viewing a product has a weight of 1.0, clicking 1.5, adding to a wishlist 2.0, adding to cart 3.0, and purchasing 5.0. Recent activities are given more importance through a time-decay method with a 45-day half-life. For new customers or users with fewer than five product interactions, the system recommends popular products as an alternative. The recommendation results are stored using Redis caching and can be accessed through the system API. The model is evaluated using temporal hold-out testing with metrics such as Precision@10, Recall@10, MAP@10, and catalog coverage. Overall, the project demonstrates how machine learning can be integrated into a practical e-commerce system to provide personalized product recommendations while supporting normal online shopping and management processes.

An Intelligent Dealer Retention and Engagement Platform Using Machine Learning, Natural Language Processing, and Explainable AI

In the highly competitive telecommunications industry, indirect sales channels managed by independent dealers contribute significantly to market share and revenue growth. However, dealer churn and declining engagement present severe operational challenges. This paper proposes a comprehensive framework for an Intelligent Dealer Retention and Engagement Platform that leverages Machine Learning (ML), Natural Language Processing (NLP), and Explainable AI (XAI) to proactively mitigate dealer attrition. The system architecture comprises four core pillars: (1) a Dealer Churn Prediction model utilizing XGBoost, optimized via F1-Score and evaluated using ROC-AUC; (2) an Explainable AI layer driven by SHAP (SHapley Additive exPlanations) to provide local and global feature transparency; (3) a Dealer Feedback Classification engine using TF-IDF and Logistic Regression to process qualitative field complaints; and (4) an automated Rule-Based Recommendation engine that translates predictive insights into prescriptive retention strategies. Utilizing Business Intelligence (BI) historical snapshots and qualitative agent visit data, this research demonstrates how foundational, highly interpretable ML models can be synthesized into a robust enterprise solution. The proposed framework bridges the gap between predictive accuracy and operational decision-making, offering telecom operators a practical blueprint for sustainable channel partner management.

Machine Learning-Based Loan Risk Prediction for Credit Scoring

The rapid expansion of consumer lending, digital banking, and other technology-enabled financial services increases the need for timely and consistent credit-risk assessment. Traditional loan underwriting commonly relies on predefined rules and manual review of income, employment, collateral, guarantors, and credit-bureau information. Although such approaches remain important, they can be slow and may have limited ability to identify complex, non-linear relationships among borrower and loan characteristics. This study develops and evaluates a machine-learning-based loan risk prediction framework for binary classification of loans into lower-risk and higher-risk outcomes. The project uses an approximately one-million-record historical loan dataset obtained from Kaggle, with 57 attributes in the supplied schema. The methodology includes data cleaning, missing-value handling, categorical encoding, numerical scaling, outlier treatment, feature engineering, stratified train-test splitting, model training, hyperparameter tuning, and iterative refinement. Five supervised learning approaches were compared: Logistic Regression, Decision Tree, Random Forest, Gradient Boosting, and XGBoost. The reported test ROC-AUC scores were 0.7024, 0.5515, 0.7165, 0.7193, and 0.7266, respectively, making XGBoost the best untuned model. After hyperparameter tuning, XGBoost achieved a test ROC-AUC of 0.7273. The results indicate that XGBoost provides the strongest overall discrimination among the tested models, while threshold selection remains a critical business decision because it changes the balance between approving good customers and preventing default exposure. The proposed framework also includes model export using joblib and deployment through a REST API and web application. The study demonstrates the potential of machine learning as a decision-support tool for credit risk management, while emphasizing that local Cambodian loan data, cost-sensitive thresholding, monitoring, and explainability are required before production deployment.

Credit Card Fraud Detection

Credit card fraud is a significant challenge in electronic payment systems because fraudulent transactions can cause financial losses and reduce customer trust. This project develops a machine-learning-based Credit Card Fraud Detection System to identify potentially fraudulent transactions from transaction data. The dataset contains 100,000 transaction records with attributes including transaction date, amount, merchant, transaction type, and location. The data is processed through feature engineering, categorical encoding, numerical scaling, and SMOTE to address class imbalance. Four machine learning algorithms—Logistic Regression, Decision Tree, Random Forest, and XGBoost—are trained and compared using Accuracy, Precision, Recall, F1 Score, and Confusion Matrix. F1 Score is used as the primary metric for selecting the best-performing model because it provides a balance between Precision and Recall for the imbalanced fraud detection problem. The selected model is integrated into a Streamlit web application that allows users to upload transaction data, predict fraudulent transactions, visualize prediction results, and download a report containing detected fraudulent transactions. The system demonstrates how machine learning can support automated transaction screening and provide a practical prototype for fraud detection.

Machine Learning-Based Demand Prediction for Factory Production Planning of Shampoo, Body Wash, and Laundry Detergent (Regression)

Accurate demand prediction can support factory production planning by reducing the risk of overproduction, stockouts, and inefficient inventory decisions. This project develops a supervised machine learning regression workflow for shampoo, body wash, and laundry detergent. Because real factory records were not supplied, a synthetic demonstration dataset was used transparently to implement and test the proposed pipeline without claiming real factory performance. The dataset contains 177 monthly product records and includes product type, month and season, previous sales, stock quantity, price, promotion, customer order quantity, and previous production quantity. Three regression algorithms - Linear Regression, Decision Tree Regressor, and Random Forest Regressor - were trained using a time-aware train/test split. Performance was evaluated using mean absolute error (MAE), root mean squared error (RMSE), and coefficient of determination (R2). On the synthetic test set, Linear Regression achieved the best result with MAE 62.5, RMSE 81.2, and R2 0.913. The workflow also produces feature-importance evidence and an illustrative production recommendation based on predicted demand, current stock, and a 10% safety-stock rule. The project demonstrates an end-to-end, reproducible framework that can be rerun with authorized real factory data for operational use.

មេរៀនទី ៦៖ Decorators and Cross-Cutting Concerns

មេរៀនទី ៦ ផ្ដោតលើ Decorators និង Cross-Cutting Concerns ដើម្បីឱ្យនិស្សិតយល់ពីរបៀបបន្ថែម Behavior ទៅ Function ឬ Method ដោយមិនលាយ Logic បន្ថែមចូលក្នុង Core Business Logic។ និស្សិតរៀនពី First-Class Functions, Higher-Order Functions, Closures, Wrapper Functions, *args, **kwargs និង functools.wraps ព្រមទាំង Decorators ដែលមាន Parameter។ ក្នុងការអនុវត្ត និស្សិតប្រើ Decorators សម្រាប់ Logging, Timing, Validation, Authentication, Authorization និង RBAC Permission Checking ដូចជា @require_login និង @require_permission(...) ហើយសិក្សាពីសារៈសំខាន់នៃ Decorator Order និង Separation of Concerns។ សរុបមក មេរៀននេះជួយឱ្យនិស្សិតបង្កើត Code ដែលអាចប្រើឡើងវិញបាន ងាយអាន ងាយសាកល្បង និងត្រៀមសម្រាប់ SOLID Principles និង Flask-based application design នៅមេរៀនបន្ទាប់។

មេរៀនទី ៥៖ Factory Method and Abstract Factory

មេរៀនទី ៥ ផ្ដោតលើ Factory Pattern ដែលជាគន្លឹះសម្រាប់បំបែក Object Creation ចេញពី Business Logic និង Client Code ដើម្បីកាត់បន្ថយ coupling និងធ្វើឱ្យប្រព័ន្ធងាយពង្រីក។ និស្សិតរៀនពី Simple Factory, Factory Method, @classmethod ជា Alternative Constructor, Registry-Based Factory និង Abstract Factory សម្រាប់បង្កើត Product ឬ Product Family ដែលពាក់ព័ន្ធគ្នា។ ក្នុងការអនុវត្ត និស្សិតបង្កើត NotifierFactory, UserFactory និង Report/UI Factory ដើម្បីជ្រើស Concrete Objects តាម runtime requirement ឬ configuration ហើយភ្ជាប់ជាមួយ Abstraction, Protocol និង Dependency Injection។ សរុបមក មេរៀននេះជួយឱ្យនិស្សិតយល់ថា Factory Pattern គឺជាវិធីរៀបចំ Object Creation ឱ្យមានរចនាសម្ព័ន្ធច្បាស់ ងាយសាកល្បង ងាយបន្ថែម implementation ថ្មី និងត្រៀមសម្រាប់ SOLID Principles និង Flask Application Factory នៅមេរៀនបន្ទាប់។

មេរៀនទី ៤៖ Abstraction, Composition, Mixins and Protocols

មេរៀនទី ៤ ផ្ដោតលើ Abstraction, Composition, Mixins និង Protocols ដើម្បីឱ្យនិស្សិតយល់ពីការរចនាទំនាក់ទំនងរវាង Objects និង Classes ឱ្យមានភាពបត់បែន និងងាយពង្រីក។ និស្សិតរៀនប្រើ ABC និង @abstractmethod សម្រាប់កំណត់ Contract, ប្រៀបធៀប is-a និង has-a relationships, ប្រើ Composition និង Dependency Injection ដើម្បីកាត់បន្ថយ coupling, ប្រើ Mixins សម្រាប់ចែករំលែក behavior និងយល់ពី MRO និង super() ក្នុង multiple inheritance។ លើសពីនេះ និស្សិតសិក្សា Protocol និង Structural Typing ដើម្បីអនុញ្ញាតឱ្យ Objects ផ្សេងៗសហការគ្នាតាម behavior ដោយមិនចាំបាច់ inherit ពី Base Class ដូចគ្នា។ ក្នុងការអនុវត្ត និស្សិតបង្កើត Multi-Channel Notification System ដែលមាន Email, SMS, Console និង Service Layer ដោយប្រើ Abstraction, Composition, Protocol និង Dependency Injection ជាគ្រឹះសម្រាប់មេរៀន Design Patterns បន្ទាប់។

មេរៀនទី ៣៖ Python Object Model and Special Methods

មេរៀនទី ៣ ផ្ដោតលើ Python Object Model និង Special Methods ដើម្បីឱ្យនិស្សិតយល់ពីរបៀបដែល Python គ្រប់គ្រង Object តាម Identity, Type និង Value ព្រមទាំងភាពខុសគ្នារវាង is និង ==។ និស្សិតរៀនប្រើ __str__() និង __repr__() សម្រាប់ការបង្ហាញ Object, __eq__() និង __hash__() សម្រាប់ Equality និង Hashing, __len__(), __iter__(), __contains__() និង __getitem__() សម្រាប់ Collection-Like Behavior, ព្រមទាំង Operator Overloading និង Context Manager តាមរយៈ __enter__() និង __exit__()។ ក្នុងការអនុវត្ត និស្សិតពង្រីក RBAC Model ដោយបង្កើត UserDirectory, PermissionSet និង AuditSession ដើម្បីឱ្យ Custom Objects អាចប្រើ Python Syntax បានយ៉ាងធម្មជាតិ ច្បាស់លាស់ និងមានសុវត្ថិភាព។

មេរៀនទី ២៖ Smart Class Design with Dataclass and Property

មេរៀនទី ២ ផ្ដោតលើការរចនា Class ឱ្យមានប្រសិទ្ធភាព ដោយប្រើ dataclass សម្រាប់ Data-Centric Models និង Regular Class សម្រាប់ Behavior-Centric Models។ និស្សិតរៀនប្រើ field(default_factory=...), __post_init__(), ClassVar, InitVar, frozen=True, slots=True, kw_only=True និង @property ដើម្បីគ្រប់គ្រង Object State, Validation, Safe Defaults និង Computed Values។ ក្នុងការអនុវត្ត និស្សិតបង្កើត Department, Permission, Role និង User ដោយប្រើ Composition ក្នុង RBAC Model ហើយអនុវត្ត is_admin, remove_role(), from_dict() ព្រមទាំង Valid និង Invalid Test Cases។ សរុបមក មេរៀននេះជួយឱ្យនិស្សិតយល់ថា Smart Class Design គឺធ្វើឱ្យ Object ងាយបង្កើតក្នុងស្ថានភាពត្រឹមត្រូវ ងាយអាន ងាយសាកល្បង និងពិបាកប្រើខុស។

មេរៀនទី ១៖ Advanced OOP Foundations, Project Structure និង Mini RBAC Model

មេរៀនទី ១ ផ្ដោតលើការពង្រឹងមូលដ្ឋាន Advanced OOP ដោយឱ្យនិស្សិតយល់ច្បាស់អំពី Class, Object, __init__(), self, Instance និង Class Attributes ព្រមទាំង Instance Method, Class Method និង Static Method។ និស្សិតអនុវត្តការរៀបចំ Python Project ជា Modules និង Package ដោយបង្កើត Mini RBAC Model ដែលមាន User, Role និង Permission តាមទំនាក់ទំនង Composition៖ User មាន Roles ហើយ Role មាន Permissions។ លំហាត់ជួយឱ្យនិស្សិតអនុវត្ត Email Validation, Role Assignment និង Removal, Permission Checking និងការសាកល្បងលទ្ធផលទាំង True និង False។ សរុបមក មេរៀននេះបង្រៀនឱ្យនិស្សិតបង្កើត Classes ដែលមានតួនាទីច្បាស់ រៀបចំកូដឱ្យងាយអាន ងាយសាកល្បង និងអាចពង្រីកទៅមេរៀន Advanced OOP បន្ទាប់បាន។

Property Valuation in Cambodia

Cambodia's real estate market has expanded rapidly in recent years, yet property pricing across the country remains largely informal. Valuations are commonly set through agent intuition, word-of-mouth, and social-media asking prices rather than through a consistent, evidence-based process. This informal approach produces inconsistent and sometimes inflated prices, widens the information gap between buyers and sellers, and slows mortgage and collateral assessments for banks and lenders. This paper presents a machine-learning-based Automated Valuation Model (AVM) for Cambodian residential property, developed as a final Machine Learning course project at Norton University. A dataset of 1,001 property listings, collected across Cambodian cities and districts, was cleaned and transformed through currency and numeric cleaning, text-based feature extraction, geographical parsing, categorical classification, and dimensionality reduction. After filtering price and land-size outliers and engineering a district-level price indicator, a final modelling dataset of 880 records and seven predictor features — bedrooms, bathrooms, land size, a unified building-area measure, district price level, city, and property type — was used to train a Histogram Gradient Boosting Regressor (HGBR). Numeric features were median-imputed and robust-scaled, categorical features were one-hot encoded, and the target price was log-transformed to reduce the influence of high-value luxury listings. The model was trained on 85% of the data and evaluated on a 15% held-out test set. On the test set, the model achieved an R² of 0.9373, a Mean Absolute Error of $34,617.38, a Mean Absolute Percentage Error of 6.01%, and produced predictions within ±20% of the true price for 96.97% of test listings. These results indicate that HGBR can capture the non-linear, location-dependent relationships that shape Cambodian property prices, with accuracy comparable to benchmarks reported for XGBoost and Random Forest models in other property markets. The resulting model offers a practical reference tool for buyers, sellers, agents, investors, and lenders, while its reliance on asking-price listings, a single-country dataset, and simplified outlier thresholds point toward clear directions for future refinement.