Systematically interpreting the results from Internet latency measurements from tools and datasets is a daunting task due to measurement noise. Here, noise is defined as the presence of non-representative values that are hard to discern from the actual network behavior. Consequently, noisy measurements lead to suboptimal operational changes and confound the actual network behavior. State-of-the-art denoising techniques are too simple, time-consuming, and labor-intensive. The main goal of this research is to investigate machine learning-based techniques to denoise Internet latency measurements. To this end, the major activities of this project are on four fronts: (a) research, (b) teaching, (c) student mentoring and professional development, and (d) dissemination and community outreach. 

(a) Research. We designed and developed a weak supervision-based framework to denoise network latency measurement. Next, we demonstrated the framework's efficacy in lab-based settings (e.g., on live latency measurements) and community datasets (e.g., CAIDA's Ark and RIPE's Atlas projects). Next, we deployed and tested the framework on live telemetry collected from modern programmable switches at the University of Oregon. Next, we extended the framework to consider multiple learning tasks (e.g., congestion, loss, noise removal). Finally, to facilitate independent validation of the results from this research, we have created a webpage (https://ix.cs.uoregon.edu/~ram/CRII.html) with project details and publications that contain analyses, results, and source code produced by this research.

(b) Teaching. We incorporated project-related outcomes in the form of new lectures on AI/ML for networking management, data features, and applications; and introduced students to some of the "good practices" in designing and evaluating data science systems for network management. These lectures were integrated into three courses (i.e., graduate networking course, graduate distributed systems, and senior- and graduate-level network measurements) at the University of Oregon. We also designed a new seminar on Internet data science at the University of Oregon. As part of the seminar, we reviewed and discussed research papers from top-tier networking, systems, and measurement conferences, emphasizing data science techniques. 

(c) Student mentoring and professional development. We mentored and trained two doctoral students, one master's student, and four undergraduate students on project-related activities. One of the undergraduate students later joined the PI's research group as a doctoral student. The undergraduate students earned jobs at companies, leveraging the data-driven skills gained from this project.

(d) Dissemination and community outreach. The above activities resulted in nine papers (two journals, four conferences, and three workshops), each of which are disseminated in prestigious networking and machine conferences and journals (e.g. NSDI, JSAC Machine Learning in Communications and Networks), and three theses. We also disseminated project outcomes via the RISE (netwoRkIng SystEms) summer school at the University of Oregon, and presentations at different fora (e.g. CAIDA's WOMBIR workshop, UO's Undergraduate Research Symposium).

Last Modified: 08/22/2022
Modified by: Ramakrishnan Durairajan