In this paper, we study the effects of using an algorithm-based risk assessment instrument (RAI) to support the prediction of risk of violent recidivism upon release. The instrument we used is a machine learning version of RiskCanvi used by the Justice Department of Catalonia, Spain. It was hypothesized that people can improve their performance on defining the risk of recidivism when assisted with a RAI. Also, that professionals can perform better than non-experts on the domain. Participants had to predict whether (...) a person who has been released from prison will commit a new crime leading to re-incarceration, within the next two years. This user study is done with (1) general participants from diverse backgrounds recruited through a crowdsourcing platform, (2) targeted participants who are students and practitioners of data science, criminology, or social work and professionals who work with RisCanvi. We also run focus groups with participants of the targeted study, including people who use RisCanvi in a professional capacity, to interpret the quantitative results. Among other findings, we observe that algorithmic support systematically leads to more accurate predictions from all participants, but that statistically significant gains are only seen in the performance of targeted participants with respect to that of crowdsourced participants. Among other comments, professional participants indicate that they would not foresee using a fully-automated system in criminal risk assessment, but do consider it valuable for training, standardization, and to fine-tune or double-check their predictions on particularly difficult cases. We found that the revised prediction by using a RAI increases the performance of all groups, while professionals show a better performance in general. And, a RAI can be considered for extending professional capacities and skills along their careers. (shrink)

In this paper we investigate risk prediction of criminal re-offense among juvenile defendants using general-purpose machine learning algorithms. We show that in our dataset, containing hundreds of cases, ML models achieve better predictive power than a structured professional risk assessment tool, the Structured Assessment of Violence Risk in Youth, at the expense of not satisfying relevant group fairness metrics that SAVRY does satisfy. We explore in more detail two possible causes of this algorithmic bias that are related to biases in (...) the data with respect to two protected groups, foreigners and women. In particular, we look at the differences in the prevalence of re-offense between protected groups and the influence of protected group or correlated features in the prediction. Our experiments show that both can lead to disparity between groups on the considered group fairness metrics. We observe that methods to mitigate the influence of either cause do not guarantee fair outcomes. An analysis of feature importance using LIME, a machine learning interpretability method, shows that some mitigation methods can shift the set of features that ML techniques rely on away from demographics and criminal history which are highly correlated with sensitive features. (shrink)

Faces provide information about a person's identity, as well as their sex, age, and ethnicity. People also infer social and personality traits from the face — judgments that can have important societal and personal consequences. In recent years, deep convolutional neural networks (DCNNs) have proven adept at representing the identity of a face from images that vary widely in viewpoint, illumination, expression, and appearance. These algorithms are modeled on the primate visual cortex and consist of multiple processing layers of simulated (...) neurons. Here, we examined whether a DCNN trained for face identification also retains a representation of the information in faces that supports social‐trait inferences. Participants rated male and female faces on a diverse set of 18 personality traits. Linear classifiers were trained with cross validation to predict human‐assigned trait ratings from the 512 dimensional representations of faces that emerged at the top‐layer of a DCNN trained for face identification. The network was trained with 494,414 images of 10,575 identities and consisted of seven layers and 19.8 million parameters. The top‐level DCNN features produced by the network predicted the human‐assigned social trait profiles with good accuracy. Human‐assigned ratings for the individual traits were also predicted accurately. We conclude that the face representations that emerge from DCNNs retain facial information that goes beyond the strict limits of their training. (shrink)