Clustering workflow requirements using compression dissimilarity measure

Xerox offers a bewildering array of printers and software configurations to satisfy the needs of production print shops. A configuration tool in the hands of sales analysts elicits requirements from customers and recommends a list of product configurations. This tool generates special question and answer case logs that provide useful historical data. Given the unusual semi-structured question and answer format, this data is not amenable to any standard document clustering method. We discovered that a hierarchical agglomerative approach using a compression-based dissimilarity measure (CDM) provided readily interpretable clusters. We compare this method empirically to two reasonable alternatives, latent semantic analysis and probabilistic latent semantic analysis, and conclude that CDM offers an accurate and easily implemented approach to validate and augment our configuration tool.