Evaluation of Interactive Machine Learning Systems

Abstract

The evaluation of interactive machine learning systems remains a difficult task. These systems learn from and adapt to the human, but at the same time, the human receives feedback and adapts to the system. Getting a clear understanding of these subtle mechanisms of co-operation and co-adaptation is challenging. In this chapter, we report on our experience in designing and evaluating various interactive machine learning applications from different domains. We argue for coupling two types of validation: algorithm-centered analysis, to study the computational behaviour of the system; and human-centered evaluation, to observe the utility and effectiveness of the application for end-users. We use a visual analytics application for guided search, built using an interactive evolutionary approach, as an exemplar of our work. We argue that human-centered design and evaluation complement algorithmic analysis, and can play an important role in addressing the "black-box" effect of machine learning. Finally, we discuss research opportunities that require human-computer interaction methodologies, in order to support both the visible and hidden roles that humans play in interactive machine learning.

A Survey of Human‐Centered Evaluations in Human‐Centered Machine Learning

“…Multiplicity of stakeholders poses a challenge to the assessment of validity and result generalizability. User segmentation is confounded by factors including cultural and educational background, age, gender, expertise, moral, and social contexts [BBL18]. In the context of HCML, all these factors coupled with subtle differences like personality traits can influence how an action is perceived, reacted to, and executed.…”

“…A challenge of current HCML research is the ability to provide nuanced evaluations of systems, given their complexity and mul-tifaceted nature. Most papers provided small-scale evaluations of simplified and encapsulated tasks [BBL18]. The well-established methodology for ML evaluation (e.g., accuracy, F-score, squared error) only covers some result-oriented aspects of human work, such as their impact on model quality.…”

Challenges in Evaluating Interactive Visual Machine Learning Systems

“…Humans can also act on their reactions to these models, such as by altering model parameters. This brings forth not only numerous intelligibility and usability issues, but also many open questions with respect to the evaluation of the various facets of the IVML system, both as separate components, and as a holistic entity that includes both human and machine intelligence [5,31]. For example, IVML tools need to be assessed more generally on their ability to increase task efficiency and correctness, as well as other possible metrics.…”

“…This question of how to elicit feedback from the analyst does not only pertain to measuring success. IVML systems can be trained by user interactions that are implicit or explicit [5], and it is possible we need to adapt our metrics accordingly.…”

Learning User Preferences via Reinforcement Learning with Spatial Interface Valuing

“…Implicit human feedback on the other hand, is a mechanism through which a human can guide an iML agent's learning process through subtle cues, such as body language. And yet a third type of feedback is known as mixed human feedback which combines explicit and implicit feedback [17]. Although all three forms of feedback impose some cognitive load on the human, explicit feedback carries the heaviest cognitive burden, especially in real-world settings where humans have additional cognitive loads due to the environment.…”