Abstract

This paper has explored survival analysis models and their ability to predict customer driven replace- ments of radio units using customer data from Ericsson. Survival Analysis examines how long until an event of interest occurs. While predictive performance is important, it does not provide insight as to how features affect the model predictions. Therefore, this study has also focused on interpreting feature-risk relationships using external explanation methods as well as model coefficients when pos- sible. This paper implemented four different models to compare; Weibull Accelerated Failure Time, Random Survival Forest, DeepHit and Logistic Hazard. A Kaplan-Meier model was implemented as a baseline for comparison. The data used consists of counters, alarms and restarts. The counters were sampled from a single weekday every week for 2025 while alarms and restarts were summed over each week. The models were evaluated using four different metrics: time-dependent concordance index, integrated brier score, time dynamic AUC and precisionk. All models performed substantially better than the baseline. Random Survival Forest was the best performing model in all metrics except precisionk, while Logistic Hazard was the best at precisionk. This suggests that Logistic Hazard is better at ranking short term risk while Random Survival Forest is better at long term risk ranking. The interpretability was evaluated using SHAP, LOFO and when possible the model’s coefficients. The most stable SHAP values were found for Weibull Accelerated Failure Time and decreasing in the order Random Survival Forest, DeepHit and Logistic Hazard. The models show consistency across the most important features, however they vary in how the features influence predictions. This study finds that a simpler model, such as the regression based Weibull Accelerated Failure Time model, provides better interpretability of results, with similar predictive performance. In contrast, combining multiple models and explanation methods provides a fuller understanding of underlying data patterns not captured by standard regression approaches.