StreamSoNGv2: Online Classification of Data Streams Using Growing Neural Gas

When dealing with unbounded streaming data, such as network packets or frames from a continuous live video feed, it is not feasible to apply iterative algorithms over the full dataset. The streaming soft neural gas (StreamSoNG) algorithm proposed by Wu et al. is particularly appealing given its ability to model arbitrary topologies in the data, however there are some shortcomings that make it difficult to apply in a practical setting. In this work, we propose and evaluate solutions to these shortcomings. Particularly, we offer an automated, data-driven approach to determining the value of a parameter
$\eta$
to which the original algorithm is especially sensitive. We demonstrate that StreamSoNGv2 is competitive with related algorithms, shows improvement over the original, and that it provides useful soft labels for each streaming data point.