Matrix completion of world trade: An analysis of interpretability through Shapley values

Abstract

Economic complexity and machine learning have recently become popular approaches for analysing international trade. However, for effective use of machine learning in relation to economic complexity and policymaking, it is important to understand what are the key features for predictions. In this framework, this article addresses the issue of the interpretability of results obtained with a machine learning technique—namely, matrix completion—when applied to economic complexity, specifically in predicting revealed comparative advantages (RCAs) of countries in different product categories. Shapley values are used to measure the role each country plays in predicting the RCAs of other countries. Countries relevant for prediction may differ from countries whose RCA values are similar to those of the country of interest when a standard similarity measure such as cosine similarity is used. We demonstrate the usefulness of our approach to identifying comparable countries by focussing our analysis on export diversification into complex goods of selected European countries.

The Impact of the European Union Emissions Trading System on Carbon Dioxide Emissions: A Matrix Completion Analysis

“…MC is a state-of-the-art quantitative method particularly suited for counterfactual analyses, as recently demonstrated by 4 , where it was successfully compared with other commonly-adopted econometric methods such as DiD and SCM. Other effective applications of MC were made by 42 , in which MC was exploited in the context of international trade for the reconstruction of World Input-Output Database (WIOD) subtables 43 , by 44 and 45 , in which MC was used for the analysis of economic complexity, and by 46 and 47 , where MC was exploited for job analysis.…”

“…3) For future research, we also consider developing a more sophisticated model that examines (in monetary terms) whether or not the decrease in output due to the price of CO 2 emissions outweighs the reduced environmental damage (this would require an appropriate definition of green GDP). In addition, the whole analysis could be extended to a less aggregate level (with larger matrices) after accelerating/parallelizing the MCFE implementation, as was done recently by 45 for another application of the related baseline MC (MCB) method. This is important for the potential use of MCFE with three-dimensional data set of countries, industries, and years, possibly achieved by associating country-industry pairs with a single index, to avoid replacing MC with tensor completion.…”

Identifying key products to trigger new exports: an explainable machine learning approach

“…Recently, the computer science community felt the necessity to provide tools (such as Shapley values) to increase the interpretability of machine learning models; this led to a number of theoretical results and practical investigations [27][28][29]. For an application of Shapley values to economic complexity-related issues, see [30,31].…”