Attention-Enhanced Autoencoder with Marginal-Variance-Regularized Feature Reconstruction for Imbalanced Insurance...

Identifying the small group of customers who hold a given policy in severely imbalanced tabular data is a recurring screening problem in insurance analytics. This study considers binary caravan-insurance policy-ownership classification on the COIL 2000 benchmark, where the positive-class prevalence is below 6%. The benchmark is a single cross-section, so the label describes current ownership rather than a future purchase event. We propose an Attention-based Symmetric AutoEncoder (ASAE) that combines an auxiliary symmetric reconstruction branch, a channel attention gate, and a marginal log-variance regularizer on a 32-dimensional latent representation. The regularizer operates on individual latent variances and is treated as a heuristic rather than as an estimator of joint differential entropy. Under a common tuning and evaluation protocol on a stratified partition, the ASAE is compared with five conventional machine learning methods and seven neural models. Across five paired runs, it achieves an F1-score of 0.6008 ± 0.0115 and an area under the receiver operating characteristic curve (AUC) of 0.8584 ± 0.0034. Relative to TabNet, the strongest baseline considered, the mean differences are 0.064 in F1-score (95% confidence interval 0.043-0.085) and 0.032 in AUC (95% confidence interval 0.017-0.048). The ordering is preserved across five stratified re-splits, four imbalance-handling configurations, and a complete type-consistent preprocessing rerun in which nominal attributes are one-hot encoded, oversampled with SMOTENC, and reconstructed with categorical cross-entropy losses (F1-score 0.6241 ± 0.0074, AUC 0.8702 ± 0.0050). All reported results use stratified random partitions of COIL 2000. Because 27% of the records share an identical predictor vector with another record, the official challenge separation and two grouped partitions are also defined, so that exact-duplicate and sociodemographic overlap can be isolated from the primary split. The training code, split indices, and per-run predictions used for the reported tables are publicly available.