A computational evaluation of Australia’s national economic structure has demonstrated that unsupervised machine learning methods can accurately identify and cluster key industrial sectors, corroborating and deepening traditional input–output analysis. Drawing on symmetric input–output tables from the Organisation for Economic Co-operation and Development across four benchmark years—2003, 2008, 2013, and 2018—the investigation applied principal component analysis, hierarchical clustering, and advanced network centrality metrics to trace how strategic industries drive macroeconomic interdependencies. The findings confirm that while classical linkage indicators successfully identify individual high-impact sectors, machine learning frameworks reveal the broader architecture of sectoral connectivity, showing how core industries cluster together and transmit economic influence across the domestic production system.
Input–output models have served as central instruments for macroeconomic planning and resource allocation since their inception in the 1970s. At their core, these frameworks utilize production equations and the Leontief inverse matrix to measure the direct and indirect transactions that bind economic activities together. Historically, key sector identification has relied on backward linkages, which capture an industry’s demand for intermediate inputs from the rest of the economy, and forward linkages, which reflect how an industry’s output serves as an input for subsequent production. Sectors exhibiting above-average normalized linkages are traditionally targeted for selective economic promotion because an expansion in their output is expected to stimulate widespread economic growth. However, standard linkage measures treat sectors primarily as isolated contributors, omitting the intricate cluster-level structures and mutual dependencies that modern data-driven algorithms can uncover.
To evaluate these relationships across time, the study analyzed the pure and symmetric input–output tables for Australia compiled by the OECD, which track 44 distinct industrial sectors measured at basic prices. The five-year intervals separating 2003, 2008, 2013, and 2018 provide a consistent temporal resolution to capture shifts in economic structures driven by technological progress and capital reallocation prior to the disruptions of the COVID-19 pandemic. To apply principal component analysis to this dataset, the researcher utilized an integrated input–output coefficient matrix, denoted as matrix H, defined by multiplying the direct requirements matrix A by the Leontief inverse matrix. Normalizing matrix H by its dominant eigenvalue prevents the analysis from being distorted by the raw scale of the data, down-weighting minor inter-industry transactions while emphasizing larger structural dependencies and mitigating the influence of multivariate outliers.
The methodological framework first reconciled traditional indices with algorithmic approaches by comparing normalized backward linkages with the first principal component derived from matrix H. Across the four benchmark years, the analysis revealed a strong and consistent non-linear association between normalized backward linkages and the first principal component, supported by R-squared values ranging from 41 percent to 77 percent. Non-parametric evaluation using Spearman’s rank correlation coefficient demonstrated coefficients well above 0.83 for 2003, 2013, and 2018, and approximately 0.80 for 2008. These high correlation values confirm that the primary variance captured by the dominant eigenvalue corresponds closely to systemic backward pull, substantiating the use of principal components as a mathematically sound representation for key sector identification within national accounts.
Scree plots constructed from the covariance-normalized matrix H illustrated a pronounced dominance of the first eigenvalue, followed by a steep initial drop and gradual decline across subdominant eigenvalues. Following the elbow rule for variance retention in high-dimensional economic systems, this pattern supported grouping industries primarily along the first principal component without compromising systemic explanatory power. To partition the 44 sectors into coherent groups, the study applied hierarchical agglomerative clustering and evaluated the total within-sum-of-squares criterion across candidate cluster counts. This optimization indicated that the Australian production structure was best characterized by five distinct clusters in 2003 and 2008, which consolidated into three broader clusters during the 2013 and 2018 benchmark periods.
Spatial factor maps projecting the first principal component on the horizontal axis and the second principal component on the vertical axis visually isolated the strategic sectors. Sectors occupying the North-East and South-East quadrants emerged as the primary engines of the domestic economy, distinguished by their high positive principal component values, whereas the majority of industries clustered in the Western quadrants with lower overall variance contributions. In the 2003 and 2008 assessments, sector 26, encompassing wholesale and retail trade alongside motor vehicle repair, appeared as an influential outlier situated in the South-East quadrant. In cluster theory, single outliers possessing extraordinary connectivity can be designated as standalone key sectors, reflecting their indispensable role in intermediating domestic exchange.
By 2013 and 2018, the structural topology showed increased consolidation, with a primary cluster encompassing five core industries: wholesale and retail trade (sector 26), professional, scientific, and technical activities (sector 38), administrative and support services (sector 39), financial and insurance activities (sector 36), and construction (sector 25). Dendrogram representations of these hierarchical relationships illustrated that these industries shared the longest branch lengths, mathematically denoting their superior systemic weight and deep interconnections. Furthermore, across all four benchmark periods, mining and quarrying (sector 4) and basic metals (sector 15) consistently formed an interconnected and autonomous branch, reflecting their persistent and specialized role within Australia’s resource-intensive trade and industrial matrix.
To rigorously test the reliability of these classifications, the study conducted a robustness comparison against four alternative analytical methodologies derived from graph theory and network analysis. These benchmarks included net multipliers, which measure the ratio of value-added generated across the economy relative to the final demand absorbed by a given sector, as well as eigenvector centrality, which weights an industry’s importance according to the prominence of its neighboring suppliers. The validation also incorporated authority and hub scores, which iteratively evaluate sectors based on whether they supply value to major consuming nodes or demand inputs from major supplying nodes, and PageRank vertex centrality, calibrated with a literature-standard damping factor of 0.85 to account for link distribution density across directed input–output networks.
Applying a threshold of 1.02 for total average normalized linkages to identify the top 15 sectors, the cross-method comparison demonstrated strong structural agreement across all seven evaluated models. Specifically, sectors 25, 26, 36, 38, and 39 were universally identified as top-five key economic sectors by all seven traditional, network, and machine learning techniques in 2018, as well as in prior benchmark years. Minor discrepancies emerged primarily among secondary sectors; for instance, the traditional Rasmussen method classified sectors 10, 16, 34, and 35 as key, whereas the net multiplier method excluded them, and sectors 27, 30, 40, and 42 failed to reach unity under net multiplier criteria. Nonetheless, core industries including sectors 1, 3, 4, 10, 23, 27, 30, 34, and 35 appeared consistently in the top ranks of at least four alternative methodologies, demonstrating the stability of the algorithmic identification.
These empirical findings provide valuable analytical guidance for industrial policy, capital planning, and macroeconomic risk assessment. The transition from one-dimensional linkage scores to multi-dimensional clustering confirms that modern services—such as administrative support, professional research, and finance—act as an essential infrastructural core that mutually reinforces trade and distribution activities. Simultaneously, the persistent linkage between basic metals, mining, and construction reveals a complementary industrial cluster capable of transmitting targeted economic interventions across the broader economy. By mapping both the relative rankings of individual sectors and their precise structural interdependencies within clusters, machine learning frameworks offer planners a more accurate instrument for anticipating how supply constraints, demand shocks, or selective investments propagate throughout national production networks.
The computational workflow combined two dedicated programming environments to execute the multi-stage estimations, utilizing MATLAB software to resolve the matrix equations and robustness comparisons while employing the R programming language to generate the cluster factor maps and sectoral ranking structures. In translating the national input–output system into a directed network of nodes and weighted edges, the methodological design addressed structural constraints highlighted in recent economic network literature. Because macroeconomic transaction tables exhibit an almost complete network architecture where the degrees of nodes are very close to one another, traditional graph measures like degree centrality and betweenness centrality become ineffective for differentiating sector significance. PageRank vertex centrality overcomes this limitation by scaling the influence of each inter-industry transaction by the total number of outgoing links provided by the supplying sector, ensuring that dense intermediate transactions do not obscure true systemic prominence.
The empirical findings also detailed the specific subsector composition operating within the primary service and heavy industrial clusters. Within the high-ranking administrative and support services classification, the model captured rental and leasing activities, travel agencies and tour operators, security and investigation activities, and office administrative and business support functions, all operating in close coordination with professional, scientific, and technical activities. In parallel, hierarchical dendrogram branching isolated an interdependent heavy industrial bloc consisting of construction, mining and quarrying, energy producing products, and basic metals. Because the basic metals sector supplies critical intermediate inputs across this entire grouping, strategic industrial policies that strengthen domestic metal manufacturing can initiate positive spillovers throughout mining, energy extraction, and construction activities, providing economic policymakers with clear cross-sectoral leverage points to guide targeted investment.
Subject of Research: Economic network analysis and machine learning identification of key Australian industrial sectors
Article Title: Applying machine learning to input–output analysis: a new perspective on identifying key Australian industries
Article References: Faridzad, A. (2025). Applying machine learning to input–output analysis: a new perspective on identifying key Australian industries. Journal of Economic Structures, 14(1), Article 9. https://doi.org/10.1186/s40008-025-00351-8
Image Credits: AI Generated
DOI: 10.1186/s40008-025-00351-8
Keywords: input-output analysis, machine learning, principal component analysis, hierarchical clustering, Australian economy, economic linkages, Leontief matrix, eigenvector centrality, PageRank centrality, authority and hub scores, net multipliers, OECD database, backward linkages, forward linkages, industrial policy, dendrogram analysis
Cite Scienmag News
Teresa Odom. (November 19, 2026). Machine Learning Pinpoints Australian Key Industries Across Benchmark Input-Output Tables. Scienmag. https://scienmag.com/machine-learning-pinpoints-australian-key-industries-across-benchmark-input-output-tables/
Teresa Odom. "Machine Learning Pinpoints Australian Key Industries Across Benchmark Input-Output Tables." Scienmag, 19 November 2026, https://scienmag.com/machine-learning-pinpoints-australian-key-industries-across-benchmark-input-output-tables/. Accessed 19 November 2026.
Teresa Odom. "Machine Learning Pinpoints Australian Key Industries Across Benchmark Input-Output Tables." Scienmag. November 19, 2026. https://scienmag.com/machine-learning-pinpoints-australian-key-industries-across-benchmark-input-output-tables/








