# Relative importance of temporal and location features in predicting smoking events

## TL;DR

The study finds that time-based cues are more effective than location-based cues for predicting smoking events, suggesting their use in timely smoking cessation interventions.

## Contribution

The study compares the predictive power of temporal and spatial features for smoking events using smartphone data and machine learning models.

## Key findings

- Excluding temporal features led to a substantial decrease in model performance.
- Removing spatial features had a minimal effect on prediction accuracy.
- Time-related cues are more robust and generalizable predictors of smoking behavior.

## Abstract

Pharmacological aids for smoking cessation, such as nicotine gum and lozenges, are most effective when used just before smoking triggers occur. Mobile technology can help by predicting these events and delivering timely reminders. This study examined the predictive value of temporal and spatial features available from smartphones. Thirty-eight participants self-reported 1784 smoking events during up to two weeks of ad-libitum smoking. Temporal features were extracted from timestamps, and spatial features were derived from GPS coordinates using methods such as DBSCAN, K-means, and distance-from-initial location. We trained logistic regression, random forest, and multilayer perceptron models with various half-time intervals (5–30 min). Across all modeling approaches and settings, excluding temporal features led to a substantial decrease in performance, while removing spatial features had a minimal effect. These results suggest that time-related cues are more robust and generalizable predictors of smoking behavior than location, supporting their use in just-in-time smoking cessation interventions.