On-device machine learning-based network bandwidth prediction to improve adaptive media streaming performance

Abstract

A media streaming method is disclosed in which a network environment of a sink device engaged in media streaming is estimated and at least two network throughput estimates are developed. A first network throughput estimate may be estimated from a measurement of network performance and a second network throughput estimate may be developed from a correlation of the estimated network environment to a machine learning model representing network throughput predictions. A final throughput estimate may be developed from the first and second network throughput estimates; and a representation of media content may be selected for retrieval based on the final throughput estimate. The machine learning model of network throughput may be developed over the course of prior media streaming session(s) that are performed by the sink device in which network throughput performance indicators of the streaming session(s) are stored over a predetermined interval and, upon conclusion of the interval, the model of network throughput is constructed according to a machine learning technique. Both the logging network throughput performance indicators and the building of the model of network throughput may be performed solely by the sink device, which preserves confidentiality of data representing consumer behavior during those media streaming sessions.