Explainable identification of predictive factors for successful inter-agency partnerships

The investments to achieve the Agenda 2030 for Sustainable Development exceed the political, financial, and technical capacity of individual organizations, making development partnerships essential. However, one out of five strategic partnerships between multilateral development organizations is currently unsuccessful, resulting in high costs and diminished impact. Consequently, development institutions are urged to adopt stricter selectivity in forming partnerships. This thesis presents a novel machine-learning application and a two-weight system designed to identify critical predictive factors for the success and failure of inter-agency partnerships, with the aim of merging computational methods with partnership dynamics. The research integrates text mining, sentiment, and semantic analysis, supervised machine learning with explainable artificial intelligence (XAI), and opaque algorithms to analyze 753 factors that influence partnership performance -the first such comprehensive list in the literature- and a corpus of co-financed project evaluation reports from multiple multilateral organizations, including a control group of non-co-financed projects. Four machine learning models -logistic regressions, random forests, XGBoost, and multilayered perceptron neural networks- are evaluated to distinguish between co-financed projects and non-partnered ones, as well as successful and unsuccessful projects. SHAP (SHapley Additive exPlanations) values rank the factors by criticality, identifying the most critical ten, while bootstrap sampling is used to calculate confidence intervals. XAI models identified factors such as country context, accessible complaint procedures, mutual benefit, analysis of failed partnerships, and disbursement arrangement harmonization as critical for successful partnerships. On the other hand, mismatches in the local context, lack of analysis of past partnerships, inadequate communication strategies, and the absence of scenario analysis in defining key performance indicators are associated with partnership failure. XAI techniques achieved a Matthews Correlation Coefficient (MCC) of 0.749, corresponding to a 74.9% correct prediction rate for partnership rating, whereas opaque algorithms achieved an MCC of 0.780 (78%). These results are comparable to the current 75% success rate for knowledge partnerships and 81% for project co-financing partnerships. Further improvement in the convergence between XAI and opaque algorithm outputs requires expanding the corpus. These insights offer organizations the tools to orient, semi-automate, and enhance the efficiency of the partnership appraisal process at the operational level. At the policymaking level, this research provides actionable insights and instruments for improving collaboration in multi-stakeholder environments, thereby enhancing the impact on beneficiaries and promoting efficient public expenditure. Future research may focus on refining factor formulation, deploying the models with incremental learning, and introducing causality analysis. Potential applications of this method include evaluating research proposals and predicting corporate ratings from narrative financial statements. Chapter 1 introduces the problem, reviews the state-of-the-art research, and outlines the research questions. Chapter 2 describes the scope, assumptions, and methods of this study. Chapter 3 presents the results, including ranked factors and corresponding evaluations. Chapter 4 addresses the research questions and proposes directions for future research.