Predicting Consumption of Illegal Drugs Using Demographic and Personality Features
Riya Bhaskar
23/09/2026
Drug use among UK school‑aged children is rising sharply, with more young people entering treatment, starting earlier, and using higher‑risk substances like ketamine and ecstacy. The combination of early onset, polydrug use, and widespread mental‑health difficulties shows this is a growing public‑health crisis rather than simple experimentation. It is important to reduce this as drug usage can be harmful especially to those younger. Additionally drugs nowadays are commonly mixed with other harmful substances that can cause further harm to children. I aimed to find models that showed very clearly the correlation between drug usage and personality factors so those can be used to decrease the problem. To approach this, a dataset containing 1885 samples, 12 features, and drug usage levels for 18 drugs was used. The models used were decision trees, logistic regression and KNN. Each model predicted the drug usage level from CL0 (never used) to CL6 (used in the last day) within 4 categories: legal, mildly addictive, moderately addictive and highly addictive. Most models showed a similar accuracy for each target, legal being approximately 82%, mildly addictive being roughly 60%, moderately addictive being approximately 35% and highly addictive being roughly 48%. Heatmaps were used to analyse where the model was making mistakes and it was shown through the heatmaps, though not predicting the correct value, were largely within a close range to the actual value, with exceptions for the CL0 labels and the legal substances target due to data class imbalances as it is highly unlikely for people to take hard drugs during their lifetime and many do consume legal substances such as chocolate and caffeine. It can be seen that illegal drug usage cannot be perfectly predicted with our model, indicating that we would need to have more specific factors beyond just personality and demographic traits to get a clearer understanding and increase the predictive accuracy of the models.