Mobile social media applications generate substantial volumes of encrypted network traffic and can provide valuable evidence in digital forensic investigations. Although modern encryption protocols such as Transport Layer Security (TLS) and Quick UDP Internet Connections (QUIC) protect packet payloads, observable traffic characteristics remain available for application fingerprinting. This paper introduces a publicly available benchmark dataset comprising 25,116 encrypted bidirectional network flows collected from ten widely used Android social media applications across five independent capture days. An automated processing pipeline converts raw packet captures into structured flow-level representations while excluding leakage-prone attributes. Five tree-based classifiers: CatBoost, HistGradientBoosting, LightGBM, Random Forest, and XGBoost, are evaluated using a temporally separated protocol in which Days 1–4 are used for model development and Day 5 is reserved for final testing. A validation-only ablation evaluates aggregation windows of 1, 20, 40, 60, and 80 consecutive flows, selecting and locking a 60-flow window before Day 5 evaluation. On the held-out Day 5 test set, multi-flow classification substantially improves performance over the single-flow baseline across all five classifiers. XGBoost achieves the highest Macro F1-score at the locked 60-flow configuration, reaching 0.7317±0.0077, compared with 0.3155±0.0018 under single-flow classification. Post-hoc SHAP and LIME analyses further identify encrypted fingerprints, packet-size statistics, timing characteristics, protocol information, and directional traffic behavior as important contributors to application discrimination. The resulting dataset and reproducible evaluation framework provide a contemporary benchmark for studying encrypted mobile social media traffic under temporal separation and multi-flow inference.