Fig 1.
Measuring the distance between skills and occupations.
(A) Shows a two-dimensional embedding of the 2018 skill distances, where the top 500 skills by posting frequency are visualized. Each marker represents an individual skill colored according to 13 clusters using K-Medoids clustering. As observed in the ‘Software Development’ inset, highly similar skills cluster together. Additionally, the specialized skill clusters, such as ‘Software Development’ and ‘Healthcare’ tend to lay toward the edges; whereas the more general and transferable skills lay toward the middle of the embedding and act as a ‘bridge’ to specialist skills. These individual skill distances form the basis of measuring the distance between sets of skills. (B) We leverage Skills Space to measure the distance between official Australian occupations at the 6-digit level (characterized by their skill sets) in 2018. The markers represent individual occupations, colored by technological labor automation risk, as calculated by Frey & Osborne. Occupations that require higher levels of routine, manual and/or low cognitive labor tasks tend to be at higher risk (colored darker red); whereas occupations characterised by non-routine, interpersonal, and/or high cognitive labor tasks are at lower risk (colored darker blue) over the next two decades.
Fig 2.
The Skills Space is statistically significant in representing job transitions.
(A) The x-axis shows the log-transformed Skills Space distance for a ‘True’ sample of actual transitions (shown in magenta color) and a ‘Random’ sample of simulated transitions (shown in gray color). Each Random transition is paired with an Actual transition: it shares the same ‘source’ occupation as the Actual transition but the ‘target’ occupation is randomly selected and is different to the Actual observation. The difference between the two populations is statistically significant (paired t-test, t-statistic = 4.514, p-value = 6.535 × 10−6, Cohen’s D effect size = 0.14). (B) We repeat the procedure 100 times: we generate 100 ‘Random’ populations, and we perform the statistical testing for each. The figure shows the histogram (density and rug plot) of the 100 obtained p-values, 87 of which are lower than 0.05.
Fig 3.
Validation and the Transitions Map.
Visualizes a subset of the Transitions Map, where 20 occupations and their pairwise transition probabilities can be observed. In this visualization, transitions occur from columns to rows, and dark blue depicts high transition probabilities, and white depicts low probabilities. While job transitions to the same occupation yield the highest probabilities (dark blue diagonal squares), it is clear that transitions are asymmetric. The dendrogram highlights how similar occupations cluster together, where there is a clear divide between services and manual labor occupations.
Fig 4.
The prediction accuracy scores of the different classifier model feature configurations.
The highest performance is achieved when ‘All Features’ are incorporated in the classifier models to predict occupational transitions (76%). Moreover, by incorporating all features, the standard deviation decreases (shown by the performance bars), which highlights the complementarity of the combined features and the ability to now account for the asymmetry between jobs.
Fig 5.
Here, we demonstrate the Job Transitions Recommender System using the ‘Domestic Cleaner’ occupation as an example—An occupation classified as a ‘non-essential’ occupation and that has experienced significant declines during the beginning of the COVID-19 outbreak in Australia.
(upper panel) The two-dimensional space of occupations (see Fig 1) with ‘essential’ occupations in blue markers and ‘non-essential’ occupations in red markers. (lower panel) We first use the Transitions Map to calculate the occupations with the highest transition probabilities (other than itself). These are the nodes on the right side of the flow diagram in the bottom panel of the figure, where the link colors depict posting frequency percentage change from March-April 2019 to March-April 2020. The link widths represent the posting frequency volume of March-April 2020 to indicate labor demand. The top six occupations have all experienced significant declines during the COVID-19 period; however, the seventh recommendation, ‘Aged and Disabled Carers’, is experiencing significant labor demand growth. ‘Aged and Disabled Carers’ were also classified as an ‘essential’ occupation during COVID-19 in Australia. We select this as the target occupation and then make personalized skill recommendations. We argue that workers trying to transition to another occupation should invest time and resources into skill development when (1) the skills are of high importance and (2) there is a high distance to acquire the skill. Conversely, workers should not focus on skill development if (1) the skills are low importance or (2) there is a low distance to acquire the skill.
Fig 6.
By applying Skills Space, we measure the yearly similarities between adaptive sets of AI skills against each of the 19 Australian industries from 2013–2019.
As industry skill sets become more similar to AI skills, the colored area of the radar chart expands. All industries have increased their similarity levels to AI skills, albeit at different rates. We argue that higher levels of AI similarity indicate AI skills are becoming more important to firms within an industry and that the skills gap to acquiring AI skills is narrowed. Access to these skills accelerates the rate of firms adopting AI and making productive use of the technologies, which offers a leading indicator of AI adoption and potential labor disruptions within these industries.