CaDENCE: a large-scale call disconnection event dataset from consumer Android devices

This article describes CaDENCE, a large-scale curated dataset of call disconnection events derived from pre-existing device-level telemetry collected from consumer Android smartphones. The source records were generated by an event-driven software instrumentation system integrated into the device operating system and synchronized to a manufacturer-side cloud data warehouse before the research curation process. This instrumentation used Android telephony framework APIs and manufacturer-side monitoring components to collect quality-of-service information at call termination, including signal quality, radio access technology, network context, and service-state indicators. The released dataset was built by filtering, standardizing, pseudonymizing, and labeling these telemetry records to support dropped-call characterization and related mobile network and call performance analyses. CaDENCE comprises 125,532,358 event-level records collected over 98 days, from March 9 to June 14, 2024, covering devices running Android 13 and 14. The dataset focuses on call disconnection events associated with packet-switched voice services, including Voice over Long-Term Evolution (VoLTE), Voice over New Radio (VoNR), and Voice over Wi-Fi (VoWiFi). Each record contains temporal, device, software, mobile network, radio access technology, signal quality, and service-state attributes, along with the binary label is_drop, which distinguishes dropped-call events from non-drop call termination outcomes. The data are organized as Parquet files using a Hive-style date-partitioned layout under the data_by_date/ directory and are accompanied by metadata files that describe the schema, missingness, value ranges, and daily and network-level aggregates, as well as data processing scripts. Because the public release was produced using a daily capped export strategy that prioritized dropped-call records, its class distribution is enriched for dropped-call analysis and should not be interpreted as the original call-outcome distribution in the source telemetry. CaDENCE is suitable for reuse in studies on dropped-call characterization, radio access technology transitions, device-side network diagnostics, and machine learning methods applied to event-level mobile network data.