Leveraging artificial intelligence: Harnessing big data to improve decision-making

IFAD is the only multilateral development institution exclusively focused on transforming rural economies and food systems to make them more inclusive, productive, resilient and sustainable. As our contribution to IFAD’s Innovation Challenge, we recently led a study that explored the potential of artificial intelligence (AI) to further these efforts by driving greater data-driven decision-making. Specifically, the study looked at how data science could help us strengthen the mechanisms for successful project design and improve our focus on results. This methodology, if fully implemented, could help propel IFAD as a leader in measuring and attributing development impact toward the achievement of the Sustainable Development Goals (SDGs).

Broadly, AI uses computers to automate decision-making processes. Machine learning, one of AI’s many applications and approaches, provides a collection of methods that get computers to recognize patterns in data, and then use these patterns to make predictions. This cutting-edge knowledge creation depends on “big data” – large volumes of data that are too complex for traditional processing methods, but that can be aggregated and analysed by machine learning algorithms.

To take advantage of this potential, a multi-disciplinary team of economists, data scientists and social scientists recently employed machine learning techniques to extract insights from across IFAD’s entire investment portfolio, spanning projects from 1981 to 2019. The results have been very encouraging.

This project had three main aims:

1. Use machine learning to gather a global overview of our investment distribution and outcomes, via historical portfolio classification of nearly 40 years of project implementation.
2. Extract evidence about the impact of key investments in order to enhance and accelerate our knowledge management, as well as improve the impact assessments of IFAD-supported projects in underevaluated areas.
3. Explore predictive analytics to develop algorithms that support the project cycle.

The starting point was the collection of a critical mass of both structured data, such as financial data and household-level impact assessment data, and unstructured data (e.g. text from project documentation). The team was then able to apply a mixed-methods analytic approach, through text mining, natural language processing, systematic reviews, meta-analysis, and predictive analytics, to combine both data types and uncover patterns.  

For example, text mining across more than 2,000 project documents showed that, over time, projects have increasingly reported against IFAD’s mainstreaming themes (climate change, gender, youth and nutrition), which is an indication that interventions have increasingly focused on addressing them. Likewise, further text mining of key terms related to the SDGs identified an increase in the presence of SDG-related content in project documentation for all 17 goals, with SDG2: Zero Hunger as the most frequently covered.

Additionally, in collaboration with Cornell University, we used an advanced machine learning model designed specifically for agriculture to examine project reports. The model detected the main categories of interventions undertaken by projects, the types of activities included in these interventions, and the most prevalent outcomes expected.  Overall, the model detected 2,200 different activities across project documentation. Socioeconomic interventions represented 37 per cent of the dataset, within which finance- and government-related activities were the most frequently reported. Technology interventions followed very closely (36 per cent) and comprised primarily of crop- and irrigation-focused activities.

Lastly, the third step in this project was to integrate the evidence generated through machine learning to develop models that can predict project performance and success. Predictive analytics comprises various statistical techniques taken from machine learning, data mining and predictive modelling to examine current and historical facts and make predictions about future outcomes or trends. Its purpose is to use machine-generated insights to guide discussion and critical inquiry, rather than consider them as “absolutes” used to apply to a decision-making framework.

In this case, we aimed to design algorithms to support the project cycle through ex ante predictions of performance and the likelihood of positive impacts resulting from development policies, given a specific set of portfolio and beneficiary features. To do this, the project team built two main prediction models. The first – a project-level prediction model – was conceived to identify successful portfolio performance features as an indication of which projects are likely to succeed. The team built the second – a household-level prediction model – to inform household-level targeting included in project design by determining which beneficiary and project features drive positive impact.

So far, the IFAD AI project has produced a global overview of investment types and outcomes; completed systematic reviews to document the impact of key interventions; and developed models that can predict future project-level performance and the probability of positive impacts, given certain targeting and project-level features.

This opportunity to experiment with innovative knowledge management technologies has opened up the possibility for IFAD to mainstream it into our everyday work by enabling an integrated, machine learning–driven approach to analyse project documentation and predict impacts. A comprehensive picture of our portfolio will give us the insight we need to achieve our strategic objective of making larger, fewer and more focused investments in a way that doubles our impact and sustainability in our project countries. More importantly, it shows that machine learning and other AI approaches make it possible to use existing data to explore new questions and gain the additional knowledge needed to improve the focus on results, to strengthen mechanisms for successful project design, and to improve measuring and attributing impact. Investments that are more impactful will, in turn, make even stronger contributions to reaching the SDGs.

Next steps? Refining the algorithms and cross-validating the models and results with different sets of big data. Stay tuned!

Find a copy of the report here.

 

For more information, please contact the project lead, Alessandra Garbero, Senior Economist for the Research and Impact Assessment Division (RIA).

Publication date: 03 June 2020