Software Project Risk Analysis Based On The Generalized Multiple-cause Discovery Method

Software industry is one of the mainstay industries of the global economy. However, the development of software is always companied by lots of risk factors. As so far, many software development projects are still vulnerable to failure because of the high risk. In order to reduce the risk and improve the success rate of the software projects, risk management should be necessary of the project management. Discovering and analyzing the causal relationships among the risk factors and project output is one of the most important and challenging tasks of software project risk analysis, which compared to the correlation analysis, can provide more specific knowledge support for making efficient risk planning and risk control strategies.In spite of the importance placed on causality discovery, yet few studies have been deep into the investigation of causality discovery in the field of software project risk analysis(SPRA), especially in terms of intelligent causal study, which deserves more concern and investigation. In actual practice, there could be three types of causalities, including the one-to-one, two-to-one and many-to-one causal models. The study of Hu et al published in the journal of Decision Support Systems in 2013 is the latest research of causality discovery in SPRA. This study proposes a causality discovery model based on “V structure” algorithm, but the model can only generate the two-to-one causalities. However, actually a consequence is usually caused by many factors, while the one-to-one and two-to-one causalities are the special cases. Therefore, this dissertation is designed for exploring an effective generalized multi-causality discovery method for SPRA.The Additive Noise Model(ANM) is generally regarded as an effective algorithm for discovering the direction of one-to-one causalities, but not for many-to-one causalities. As for the generalization of causal knowledge discovery, this study proposes a generalized multiple-cause discovery method(ANMCPT) by combining ANM with Conditional Probability Table(CPT), by which the issue of multi-cause inference could be converted to single-cause inference, so as to better analyze the causal relations between the risk factors and the performance of software projects. Based on the collected dataset of 498 samples, the study establishes a forecasting model using the discovered causal network structure for risk analysis of software projects. Besides, the performance of the model is tested by using the method of 10-fold cross-validation.The experiment results show that the proposed method is effective to discover one-to-one as well as many-to-one causalities in SPRA on the collected software project data. The causal knowledge discovered is in conformity with the existing expert knowledge. Besides, five subjective factors of project management risk are found as the direct and crucial causes of the project performance. Moreover, the proposed method presents high prediction accuracy and it performs better than the compared algorithms, including Na?ve Bayes, C4.5 Decision Tree, general Bayesian Networks, and regression analysis algorithms. Therefore, the causal model based on ANMCPT could be an approach for risk forecasting.As a whole, our proposed causal discovery method can be an effective strategy for risk analysis and risk prediction of software projects. This study takes an exploratory research applying the Additive Noise Model to SPRA and proposes a considerable method for discovering many-to-one causalities among the risk factors and project performance. Furthermore, the causal knowledge discovered to a certain extent could provide valuable references and guidance for the practice of risk control decision making. Thus, the conducted research somewhat has significant implications on both theory and practice.