Academic · Team of 4Case study 03 / 05

TrendRadar

Early detection of viral trends

A machine-learning system that predicts which keywords will go viral, from their very first minutes of activity.

Context: Team project · 4 members
Year: 2026
Models: XGBoost · River ARF
Learning: Offline + online

ROC-AUC: 0.921; How well the model ranks keywords that went viral above those that didn’t. 0.5 is chance; 1.0 is perfect.
Early signal: 15 min; All the model sees: a keyword’s first 15 minutes of activity.
Prediction horizon: 12 h; What it predicts: the keyword’s virality over the next 12 hours.
Tweets: ~179k; Tweets the XGBoost model was trained on.

01Context

A team project from the AI & Data Science curriculum focused on predictive modelling and online learning. TrendRadar watches the first minutes of a keyword’s activity on social platforms and estimates how viral it will become.

02Problem

By the time a trend is obviously viral, it’s too late to act on it. And social media behaviour keeps shifting, so a model trained once slowly goes stale.

03Architecture

01Sources
Social connectors
Bluesky · Mastodon
02Window
First 15 minutes
Early activity signal
03Features
Extraction + normalisation
04Model
XGBoost
Tuned with Optuna
ROC-AUC 0.921
05Output
12-hour virality
06Online
River Adaptive Random Forest
06Drift
Concept-drift detection

From the first minutes of activity to a prediction that keeps learning.

04Team

Team project (4 members) focused on early viral-trend prediction from social-media activity, combining XGBoost, feature engineering, hyperparameter optimisation, and online learning with concept-drift detection.

05Technical decisions

01
Predict from the earliest signal
Predictions come from the opening minutes of activity, so they arrive while a trend is still early enough to act on.
02
Gradient-boosted trees, tuned with Optuna
XGBoost handles the offline prediction; Optuna searches hyperparameters and feature configurations.
03
Keep learning from new data
A River Adaptive Random Forest learns online from new data and detects concept drift when patterns change.
04
Pluggable data connectors
Bluesky and Mastodon connectors are built on a shared base connector, keeping data collection separate from the models.

06Results

Virality predicted from a keyword’s opening minutes, while the trend is still early
An online model that adapts as social media behaviour shifts
A Streamlit application on top of the models