Analyzing Server Connection Patterns Using Clustering and Regression

This project was completed as part of the UBC DSCI 100: Introduction to Data Science course.
It explores how server resources can be optimized based on patterns in player connection behavior using clustering, visualization, and K-Nearest Neighbors (KNN) regression.

📁 Files

• final.ipynb: Main notebook containing all code, visualizations, and modeling.
• players.csv: Metadata of 196 players, including experience, subscription status, and gameplay hours.
• sessions.csv: 1,535 session records, including session times and durations.

🎯 Project Goal

To predict and optimize server resource allocation by analyzing when players are most active.
This involves understanding the number of concurrent connections across days and hours.

🔍 Key Steps

1. Data Cleaning and Feature Engineering

• Converted time strings to datetime objects
• Calculated session durations
• Extracted hour and day of week features
• Removed sessions with 0 duration

2. Exploratory Data Analysis (EDA)

• Plotted connections per hour and per day using Altair
• Created bubble charts to visualize activity heatmaps
• Identified late-night and weekend peak usage

3. Clustering with K-Means

• Used elbow method to determine optimal k = 3
• Grouped time slots into Low, Medium, and High connection density
• Labeled data with cluster names for better interpretability

4. Modeling with KNN Regression

Evaluated three strategies using Root Mean Squared Percentage Error (RMSPE):

Model	Average RMSPE	Notes
Full dataset (no clusters)	0.42	Least accurate
Full dataset + density labels	0.23	Most accurate
Cluster-specific models	0.32	Intermediate accuracy, more modular

5. 3D Visualization

• Created interactive 3D Plotly surfaces to show connection density across hours and days
• Compared predicted vs actual patterns using KNN-regressed surfaces

📊 Key Findings

• Peak activity between 11:30 PM and 4:30 AM, especially on Saturday 2:00 AM
• Lowest activity during weekday mornings
• Friday showed surprisingly low usage
• Density-aware KNN models performed significantly better than unclustered models

🛠️ Tools Used

• Python (Jupyter Notebook)
• pandas, numpy
• altair, plotly
• scikit-learn (KMeans, KNN, GridSearchCV)

📌 Authors

JunHyun Kim
David Liu
Layni Janzen
Sydney Lee

This project was completed as part of UBC's DSCI 100 (Winter 2024) — Final Group Project (Group 10).