Reliable internet traffic prediction can help network teams plan capacity, anticipate demand, and respond more proactively to changing usage. The same principle applies across industrial operations: historical data from a network or other infrastructure can reveal patterns that support better planning, provided the model’s limits are understood. A study of daily internet traffic at a Nigerian university illustrates how supervised machine learning can classify expected traffic levels using recent network history and calendar information.
Why network traffic forecasting matters
Network traffic monitoring produces more than a record of data volumes. Logged measurements can help teams identify usage patterns, distinguish busier from quieter periods, and assess whether a network may face rising demand. Those insights can inform decisions about network management and potential capacity needs. Traffic analysis can also support other objectives, including classifying traffic and identifying activity relevant to network security.
Making use of this information can be challenging when monitoring generates large datasets. Data mining offers one way to extract patterns from historical records by applying learning algorithms. Rather than relying only on manual inspection, a trained model can use past observations to classify a new day’s traffic into defined categories.
How the internet traffic prediction model worked
The study analysed traffic recorded at Covenant University in Nigeria from January to December 2017, covering 353 days. It treated upload and download volumes separately and classified each day’s traffic into four status categories: low, slight, moderate, or heavy. The categories were defined from the distribution of observed traffic values, using quartile-based thresholds.
For each prediction, the models used calendar details and recent traffic history. Inputs included the month, week of the year, day of the week, traffic volume from the previous day, and the average volume for the two preceding days. The target was the traffic-status classification for either daily upload or download traffic.
Before model training, the dataset was cleaned and prepared, and numerical inputs were normalized. The researchers used 70% of the observations to train the models and the remaining 30% to evaluate them. This separation is important: performance on the evaluation data gives a more useful indication of how a model handles observations it did not use during training.
Comparing algorithms and evaluation platforms
The analysis used two visual data-mining platforms, KNIME and Orange. KNIME models included Tree Ensemble, Decision Tree, Random Forest, and Naïve Bayes algorithms. Orange models included k-nearest neighbours (kNN), Random Forest, Neural Network, Naïve Bayes, and the CN2 Rule Inducer.
Performance was assessed with measures including accuracy, Cohen’s Kappa, precision, recall, F-measure, and area under the receiver operating characteristic curve (AUC). These measures describe different aspects of model behaviour. Accuracy reports the proportion of correct classifications, while precision and recall provide class-specific views of false and missed classifications. AUC evaluates how well a model distinguishes between classes across classification thresholds.
The results varied by both platform and traffic direction:
- On KNIME: Tree Ensemble had the highest download accuracy at 62.264%. For upload traffic, Random Forest performed best at 63.208%. Decision Tree had the lowest accuracy for both directions, at 55.66%.
- On Orange: Naïve Bayes had the highest AUC for both download and upload traffic. The highest classification accuracy was 58.8% for download traffic with Naïve Bayes and 63.9% for upload traffic with Naïve Bayes.
- Across the Orange models: download accuracy ranged from 51.4% for kNN to 58.8% for Naïve Bayes; upload accuracy ranged from 57.3% for kNN to 63.9% for Naïve Bayes.
These figures describe this particular dataset and classification task. They are not a general guarantee of forecast performance for other networks. The study’s evaluation also shows why it is useful to examine more than one metric: a single overall accuracy value does not reveal how reliably a model identifies each traffic category.
Practical lessons for network and operations teams
Build a usable history of measurements
Prediction depends on consistent historical data. The case study relied on daily upload and download measurements, alongside dates and recent traffic values. For an operations team, the practical starting point is to define what is measured, how it is recorded, and which time interval is relevant to planning.
Match categories to the decision
The four traffic classes made the task a classification problem: the models predicted a status category rather than an exact daily volume. That distinction matters. A category-based forecast may support broad preparation for heavier demand, while a decision requiring precise volume estimates would need an appropriately defined prediction target and evaluation approach.
Validate before using predictions operationally
Model performance was moderate, with results differing across algorithms, platforms, and upload versus download traffic. Teams should therefore evaluate performance on data held back from training and review errors by class, not just overall accuracy. A model that confuses moderate and heavy periods may have different operational value from one that performs consistently across all categories.
Visual workflow tools such as KNIME and Orange can make it possible to assemble and compare data-mining pipelines without relying exclusively on extensive custom coding. However, ease of building a workflow does not remove the need for sound data preparation, suitable evaluation, and careful interpretation.
Conclusion
Historical traffic records can provide a foundation for network traffic forecasting and more informed operational planning. The university case study demonstrates a practical classification workflow using calendar features and recent traffic measurements, while its results underscore that performance depends on the algorithm, platform, and traffic direction. For network engineers and operations leaders, the key implication is to treat machine-learning predictions as decision-support inputs: establish reliable data, select metrics that reflect operational needs, and validate results before using them to guide capacity or service planning.
Access Official Technical PDF