Psychometrics and AI

Professor John Rust, UK Test Users Conference, December 2005

Psychometrics and Artificial Intelligence

Psychometric tests and interviews are both complementary and competitive as techniques for job selection. However, while psychometrics have become increasingly sophisticated in its scientific refinement, the interview still represents, for most human resource professionals, the ideal technique if both time and the necessary training allow. However, with the advent of information technology, the distinction between the two approaches is becoming increasingly blurred. Artificial intelligence techniques, such as expert systems and neural programming are increasingly finding new uses and applications in the field, and have the potential to mimic many of the perceived benefits of the interview.

Classical psychometric tests function by summing the scores from pre-designated items within the test to form scales that represent particular traits of ability or personality. The interactions between these items and the patterns they form are ignored. It is the insight of human interviewers into the complexities of interviewees' responses that has always given them the edge over this rather simplistic approach. However, using neural network technology it is now possible to train artificial intelligence systems embodied within computers to recognise the complex patterns that exist among and between the responses to items in questionnaires. These could represent crucial aspects of human personality if only they could be made available to human resource professionals.

Not all psychometric problems can be addressed in this way, and the technique is very specific in its requirement for training data that includes information on 'successes' and 'failures' from previous administrations of the questionnaire. However, where such information is available, the neural network is able to interpret the complexity inherent in the system in a manner that eludes the classical statistical approach. For example, it is possible to train a network to recognise credit risk from a database of credit applications. Similar techniques can be applied to insurance risk and fraud detection. Indeed any structured database, whether it be of application forms, biodata or personality questionnaires can provide the raw material for training such a network.

Psychometric neural networks, by recognising the patterns of responses characteristic of attempts to cheat in a test, can operate as superior lie detectors that can be of considerable help to administrators in interpreting test results. Another application would be the training of a network to discriminate between the response patterns of 'stayers' and 'leavers' in, say, the six months following appointment, enabling data mined at the time of application to contribute towards a company strategy to reduce staff turnover.

One reason for optimism lies in the ability of the neural network to recognise different pathways to the same end. Traditional tests under-perform in this respect, and it is a cliché that to depend solely on such tests would result in teams of very similar "clone" workers, lacking the necessary diversity for effective team development. High performance in management, for example, is not the result of a single trait or indeed of a particular personality and skills profile. Rather, each high performer realises their potential in a different way, utilising their particular strengths to achieve their objectives in a pattern that is unique to them. By focussing on the ends, rather than the means, the neural network is able to recognise patterns that lie within the inherent complexity of the human character, beyond the level of the simple trait.