User Modelling for Avoiding Overfitting in Interactive Knowledge Elicitation for Prediction

Abstract

In human-in-the-loop machine learning, the user provides information beyond that in the training data. Many algorithms and user interfaces have been designed to optimize and facilitate this human-machine interaction; however, fewer studies have addressed the potential defects the designs can cause. Effective interaction often requires exposing the user to the training data or its statistics. The design of the system is then critical, as this can lead to double use of data and overfitting, if the user reinforces noisy patterns in the data. We propose a user modelling methodology, by assuming simple rational behaviour, to correct the problem. We show, in a user study with 48 participants, that the method improves predictive performance in a sparse linear regression sentiment analysis task, where graded user knowledge on feature relevance is elicited. We believe that the key idea of inferring user knowledge with probabilistic user models has general applicability in guarding against overfitting and improving interactive machine learning. * This is the pre-print version. The paper is published in the proceedings of IUI 2018 conference. Definitive

Knowledge Graph Completion-based Question Selection for Acquiring Domain Knowledge through Dialogues

“…This study is positioned at the intersection of natural language dialogue [21,40,45] and interactive machine learning [1,3,18,22,26], in which the outputs of a machine learning model are visually displayed to the user [14] or are used to select informative questions for the user [13]. Our study is close to the latter and shows a promising way to obtain knowledge from the user during dialogues.…”

Approximate Bayesian Computation with Domain Expert in the Loop

“…However, that runs the risk that the users may amplify noise in the statistics, especially in low-simulation regimes, if they are not careful. Using so-called "posterior elicitation" and inferring the priors indirectly may then be helpful (Daee et al, 2018).…”

A Survey of Domain Knowledge Elicitation in Applied Machine Learning

“…Such interfaces can be used for tasks that go beyond labeling to include relevant tasks such as feature creation, re-weighting features, adjusting cost matrices, or otherwise modifying model parameters. Others have pointed to specific threats in interactive ML related to the information a human provides, including that a user's input reinforces noise in the training data or statistics they see [52]. Recent works have also surveyed research in human-in-the-loop machine learning [53,54].…”

“…Finally, related to providing context, some recent interactive machine learning research argues that, when experts are shown labeled data used for training or validation during elicitation, the knowledge that is elicited from them can be redundant in ways that hamper model performance [52]. This possibility is largely not mentioned by papers in our sample that explicitly described giving domain experts access to training data, but as elicitation interfaces become a more focal aspect of applied machine learning research, this risk may be important for researchers to account for.…”