The Machine-Learning Predictive Recertification Targeting system is a research prototype developed by Dario Sansone of the University of Exeter Business School and Anna Zhu of RMIT University, using Australian government administrative data to predict the intensity and duration of income support receipt among welfare enrollees in the Centrelink social security system. The system was designed to forecast the proportion of time each individual would remain on income support over a subsequent four-year horizon, with the explicit aim of identifying individuals at highest risk of long-term welfare dependency so that early intervention programmes and recertification review processes could be targeted more effectively (Sansone and Zhu, 2021, IZA DP 14377, p. 1).
The research uses the DOMINO (Data Over Multiple Individual Occurrences) longitudinal administrative dataset, which is maintained by the Australian Department of Social Services and captures individuals' interactions with the welfare system without identifiable information such as names and addresses (DSS Aristotle Metadata Registry). DOMINO contains daily-frequency records of income support receipt status from 2000 onwards, covering over 32 million persons who had any contact with the Centrelink system during that period (Sansone and Zhu, 2021, p. 5). The data are high quality because the government relies on this exact information to determine eligibility for payments: an individual's payment amount is a direct function of their income, wealth, savings, household structure, and other socio-economic factors, and these data are reconciled with Australian Tax Office records to ensure accuracy (Sansone and Zhu, 2021, p. 3). Recipients' eligibility for payments is assessed regularly, and recipients are required to report changes such as to relationship status, earnings, or living conditions within 14 days of the change (Sansone and Zhu, 2021, p. 10). The dataset includes information on demographics (sex, age, country of birth, and Indigenous status), household structure, government benefit receipt history by type, personal relationships, employment and underemployment, work instability, location and residential mobility, housing, education, income, and wealth (Sansone and Zhu, 2021, p. 13). In total, approximately 1,800 possible predictive features were constructed from these administrative records (Sansone and Zhu, 2021, p. 15).
The research was funded through Australian Research Council Linkage Project LP170100472 (Sansone and Zhu, 2021, acknowledgements footnote, p. 3). The analytical sample covers the period 2014 to 2018, using 2014 as the base year for predictive features and measuring welfare receipt intensity from 2015 to 2018. A 1% random sample of approximately 50,615 individuals aged 15 to 66 was drawn from the full population for computational reasons (Sansone and Zhu, 2021, p. 10-11).
The technical approach uses an ensemble of off-the-shelf classical machine-learning algorithms: LASSO (a regularised regression method), Support Vector Regression, and Boosting (gradient-boosted trees allowing up to 6-way interactions between input variables). The data were split into an 80% training sample and a 20% hold-out test sample for out-of-sample performance evaluation (Sansone and Zhu, 2021, pp. 15-16). The ensemble method, which combines predictions from all three algorithms using weighted linear regression, achieved the best performance overall (Sansone and Zhu, 2021, p. 18).
In terms of performance, the machine-learning ensemble achieved an out-of-sample R-squared exceeding 76%, representing at least a 22% improvement (approximately 14-percentage-point increase in R-squared) compared to the best-performing OLS heuristic model and standard early warning systems currently in use (Sansone and Zhu, 2021, p. 18; University of Exeter, 2021). The authors conducted back-of-the-envelope calculations showing that individuals identified by the ML model as long-term recipients accrued an additional welfare cost of approximately AUD 0.99 billion compared with comparably sized groups identified under the existing actuarial profiling approach used in the government's Try, Test and Learn programme, representing roughly 10% of total annual unemployment benefit expenditure (Sansone and Zhu, 2021, p. 18). The ML algorithms also identified new powerful predictors not commonly associated with long-term welfare receipt, including annual income variability, residential relocation frequency, and failure to meet mutual obligation criteria (Austaxpolicy, 2021).
The relevance to recertification and exit decisions lies in the system's ability to predict which individuals are most likely to remain on income support for extended periods, thereby enabling targeted recertification scheduling and resource allocation for exit-focused interventions. The paper explicitly notes that Australia's income support payments are strictly means-tested with regular eligibility assessment, and that recipients who fail to comply with mutual obligation requirements such as activity tests and job search can face sanctions including loss of payments (Sansone and Zhu, 2021, pp. 7-9). The ML predictions could inform which recipients receive more intensive casework review and which can be managed with lighter-touch recertification processes.
The human oversight model envisaged by the researchers is explicitly advisory and complementary to caseworker expertise. The authors state that the algorithms should not replace human expertise but rather act as its complement, allowing caseworkers to focus their attention and time providing personalised service and targeting appropriate support to individuals that the algorithm identifies as most at risk (University of Exeter, 2021; IZA Newsroom, 2021). The authors also advocate for a system to monitor and audit automated decision-making, referencing the Australian Robodebt scandal as a cautionary example of the potential harms from automated welfare systems (Austaxpolicy, 2021). The predictive models can reduce conscious and unconscious biases common in human decision-making by avoiding arbitrary selection of predictors or subgroups, and have the potential to prevent cream-skimming practices where employment service providers target individuals with easier-to-achieve outcomes (Sansone and Zhu, 2021, p. 6).
The authors acknowledge limitations of the predictive approach: prediction is only a first step, and policymakers additionally require evidence on the effectiveness of specific interventions, which can only be obtained through causal methodologies such as randomised controlled trials rather than predictive modelling alone (IZA Newsroom, 2021; Sansone and Zhu, 2021, p. 7). Furthermore, the ML algorithms would need to be retrained using data from economic downturns to ensure continued accuracy during recessionary periods (Sansone and Zhu, 2021, p. 25). The authors also note persistent scepticism regarding accuracy concerns and bias reinforcement in algorithmic systems (IZA Newsroom, 2021). As of the most recent verification, this system remains a research prototype and has not been operationally deployed within Services Australia or any other Australian government agency.