| Popis: |
In principle, neural networks and other algorithms of machine learning can perfectly map unique relations of any complexity between input and output variables. We investigate whether using multi-layer neural networks allows improving personality assessments by constructing short-tests that are more efficient. Personality data for N = 3,498 participants from Germany, the US and the UK was collected using the International Personality Item Pool-300-item-version (IPIP-300 or IPIP-NEO), the Big Five Inventory (BFI-10) and the HEXACO Personality Inventory-60 (HEXACO-PI-R). We trained 40 multi-layer neural networks on this data to predict individuals’ scores on the Big-5-personality dimensions as well as facet scores from a 30-item version of the IPIP. A neural network based short-test version, IPIP30-NNet, predicted Big-5 dimensions from IPIP-300 as well as its facets with high accuracy. The correlations with the long-test scores (IPIP-300) were significantly higher compared to short-tests using a standard averaging algorithm and a multiple regression. Particularly for the facet scores, IPIP30-NNet lead to substantial improvements in predictive validity (Δr = .04 - .17). Additionally, as a syntheses of all three personality tests we calculated Big-5-“superscores”, which could be predicted from IPIP30-NNet with high accuracy as well. Our results demonstrate that neural network based diagnostic can be used to receive a very detailed individual personality profile based on very few information. We discuss challenges, potentials and future directions for using machine learning to improve standard psychological assessment. |
