A Non-linear Distance Metric for Instance-based Learning

P. Juell and P. Paulson (USA)


Neural networks, Machine-learning, Case-based Reasoning, Nearest-neighbor Classification


Proposes a new similarity metric for use in instance-based learning algorithms. An artificial neural network is trained using back-propagation to return the distance of instances from a training population based on whether the instances share the same classification. Empirical results compare a nearest-neighbor classifier using the metric to classifiers using existing metrics. The metric allows the instance based learning algorithm to have better performance on problem sets that require context. Results show that the classifier outperforms standard back-propagation classification on some data sets. The new metric requires fewer parameters than the other metrics examined.

Important Links:

Go Back