Globally, two billion people and more than half of the poorest adults do not use formal financial services. Consequently, there is increased emphasis on developing financial technology that can facilitate access to financial products for the unbanked. In this regard, smartphone-based microlending has emerged as a potential solution to enhance financial inclusion. We propose a methodology to improve the predictive performance of credit scoring models used by these applications. Our approach is composed of several steps, where we mostly focus on engineering appropriate features from the user data. Thereby, we construct pseudo-social networks to identify similar people and combine complex network analysis with representation learning. Subsequently we build credit scoring models using advanced machine learning techniques with the goal of obtaining the most accurate credit scores, while also taking into consideration ethical and privacy regulations to avoid unfair discrimination. A successful deployment of our proposed methodology could improve the performance of microlending smartphone applications and help enhance financial wellbeing worldwide.