Visual Assessment of Web User Interfaces for Interactive Machine Learning Tools
Living in an era where machine Learning (ML) and Artificial Intelligence (AI) form the very pillars of the technology industry, many companies are developing tools that can perform intelligent tasks like forecasting and critical analysis, automatically, by implementing algorithms that learns from data. On the other hand, UX/UI designers and developers are finding it difficult to design web user interfaces (WUIs) for these tools because of their highly technical and intricate structure. Even understanding and leaning machine learning’s concepts have not shown to benefit designers envision what the product might look like. Significant brainstorming effort is required to identify the best and optimized ML tool WUI design and problems directly translate into increased numbers of tickets.
Thus, determining the best approach to represent datasets, parameters or processed results requires a concrete solution that can predict the usability-, complexity- and aesthetics-related attributes of ML tools’ web user interfaces. This thesis investigates visual analysis techniques for evaluating the visual complexity and aesthetics of ML tools’ WUIs, using combinations of different metrics. The resulting analysis solution must predict objective and subjective attributes of the WUIs such as interaction times, cognitive complexity, or Kansei-scaled aesthetics. A ground truth of target values is created from UAT tests with test subjects on variations of existing ML tool WUIs, and the impact of various computable WUI metrics on those values is analyzed to create or train a suitable prediction model. This approach requires the implementation of the visual analyzer for the input metrics, planning and conducting the UAT tests for creation of the ground truth and validating the resulting model in a suitable evaluation experiment on a set of ML tool WUIs previously not used.
The objective of this thesis is to find an approach or a combination of approaches to solve the above problem of predicting usability-, complexity- and aesthetics-related attributes of WUIs in the context of ML tools based on visual UI analysis. This particularly includes the state of the art regarding visual UI analysis, metrics, and predictions. The demonstration of feasibility with an implementation prototype of the concept is part of this thesis as well as a suitable evaluation on real ML tool WUIs as detailed above.