Duplicate detection, which is an important subtask of data cleaning, is the task of identifying multiple representations of a same real-world object. Numerous approaches both for relational and XML data exist. Their goals are either on improving the quality of the detected duplicates (effectiveness) or on saving computation time (efficiency). In particular for the first goal, the "goodness" of an approach is usually evaluated based on experimental studies. Although some methods and data sets have gained popularity, it is still difficult to compare different approaches or to assess the quality of one own´s approach. This difficulty of comparison is mainly due to lack of documentation of algorithms and the data, software and hardware used and/or limited resources not allowing to rebuild systems described by others.
In this paper, we propose a benchmark for duplicate detection, specialized to XML, which can be part of a broader duplicate detection or even data cleansing benchmark. We discuss all necessary elements to make up a benchmark: Data provisioning, clearly defined operations (the benchmark workload), and metrics to evaluate the quality. The proposed benchmark is a step forward to representative comparisons of duplicate detection algorithms. We note that this benchmark is yet to be implemented and this paper is meant to be a starting point for discussion. [more]