Biblio

The development process for new algorithms or data structures often begins with the analysis of benchmark results to identify the drawbacks of already existing implementations. Furthermore it ends with the comparison of old and new implementations by using one or more well established benchmark. But how relevant, reproducible, fair, verifiable and usable those benchmarks may be, they have certain drawbacks. On the one hand a new implementation may be biased to provide good results for a specific benchmark. On the other hand benchmarks are very general and often fail to identify the worst and best cases of a specific implementation. In this paper we present a new approach for the comparison of algorithms and data structures on the implementation level using code coverage. Our approach uses model checking and multi-objective evolutionary algorithms to create test cases with a high code coverage. It then executes each of the given implementations with each of the test cases in order to calculate a cross coverage. Using this it calculates a combined coverage and weighted performance where implementations, which are not fully covered by the test cases of the other implementations, are punished. These metrics can be used to compare the performance of several implementations on a much deeper level than traditional benchmarks and they incorporate worst, best and average cases in an equal manner. We demonstrate this approach by two example sets of algorithms and outline the next research steps required in this context along with the greatest risks and challenges.

DeepQA is a large-scale natural language processing (NLP) question-and-answer system that responds across a breadth of structured and unstructured data, from hundreds of analytics that are combined with over 50 models, trained through machine learning. After the 2011 historic milestone of defeating the two best human players in the Jeopardy! game show, the technology behind IBM Watson, DeepQA, is undergoing gamification into real-world business problems. Gamifying a business domain for Watson is a composite of functional, content, and training adaptation for nongame play. During domain gamification for medical, financial, government, or any other business, each system change affects the machine-learning process. As opposed to the original Watson Jeopardy!, whose class distribution of positive-to-negative labels is 1:100, in adaptation the computed training instances, question-and-answer pairs transformed into true-false labels, result in a very low positive-to-negative ratio of 1:100 000. Such initial extreme class imbalance during domain gamification poses a big challenge for the Watson machine-learning pipelines. The combination of ingested corpus sets, question-and-answer pairs, configuration settings, and NLP algorithms contribute toward the challenging data state. We propose several data engineering techniques, such as answer key vetting and expansion, source ingestion, oversampling classes, and question set modifications to increase the computed true labels. In addition, algorithm engineering, such as an implementation of the Newton-Raphson logistic regression with a regularization term, relaxes the constraints of class imbalance during training adaptation. We conclude by empirically demonstrating that data and algorithm engineering are complementary and indispensable to overcome the challenges in this first Watson gamification for real-world business problems.