Gait authentication using wearable sensors has attracted attention as a promising approach for continuous authentication. However, in many existing studies, authentication models were trained only on short-term gait data collected under controlled experimental conditions, which may not reflect the diversity of gait patterns in daily life. In this study, we investigated the effect of incorporating long-term gait data collected in daily-life environments into the training dataset of a deep-learning-based gait authentication model using an insole-type gait sensor. Gait data were collected from 48 healthy participants using two measurement modes: a real-time mode for short-term data and a daily-life mode for long-term data. Two training conditions were compared: using only short-term data and using both short-term and long-term data. Authentication performance was evaluated for unseen users using the equal error rate (EER). Incorporating long-term data tended to improve the performance, reducing the average EER from 0.47 to 0.17% and the common-threshold EER from 0.71 to 0.28%. In addition, feature-space analysis using the Silhouette Score and Delta Cosine Similarity (ΔCosSim) showed improved separability of test embeddings, and SHapley Additive exPlanations (SHAP) based analysis revealed reduced dependence on roll-angle-related parameters and increased use of toe-in/out angle and minimum toe clearance.