Towards an ML Assisted DASH-based Architecture: Leveraging Predictive Network Analyses with Interpretability

Eduardo R. Peretto, Manoel Narciso Reis Soares Filho, Débora Cristina Santos Sousa, Luciano Paschoal Gaspary, Bruno Iochins Grisci

Abstract: While Dynamic Adaptive Streaming over HTTP (DASH) is the standard for media delivery, most adaptive bitrate (ABR) algorithms remain reactive. Integrating predictive insights without compromising system design is a key challenge. This paper presents a feasibility study of an ML-enhanced DASH architecture that generates lightweight, interpretable prediction hints to enable proactive ABR. We validate the framework using a case study on over 10,000 hours of real traffic traces from Brazil’s Rede Ipê backbone. Using Random Forest and Gradient Boosting models, we compare a DASH-only feature set against one enriched with network metrics (RTT, traceroute). Our results demonstrate the viability of the architecture and highlight key network indicators that drive predictions. By focusing on interpretability and statistical validation, our work provides a transparent framework for integrating predictive modules into DASH ecosystems, laying the groundwork for more robust, nextgeneration ABR algorithms.

External IDs:dblp:conf/cnsm/PerettoFSGG25