Scholarlink

Test-driven development has obtained nearly universal adoption in back-end programming, and now front-end testing automation sees extensive progress. To advance beyond auto-checking the purely “syntactic” correction of code and designs of user interfaces (UIs), it is necessary to predict quantifiable dimensions of user satisfaction and experience (UX) – particularly, visual aesthetics. In this task, neural networks (NNs) already largely overpower the approaches based on cognitive architectures that were popular in the past. However, datasets in Human-Computer Interaction (HCI) generally have limited sizes, while pre-trained models for predicting aesthetic impressions do not transfer well between different website domains and user groups. In our paper, we employ openly available datasets by Calista (overall 42,000 aesthetic evaluations for about 500 website screenshots) and compare available NN models (AlexNet, CaffeNet, and NIMA-MobileNet) with varying numbers of layers and the representation of the output. Fine-tuning AlexNet with another open dataset, Web-AVA, yielded the best improvement of RMSE by about 5%. The results of our study might be of interest for both AI/ML researchers and practitioners, and for web designers who seek to catch up with the today’s NN-based testing automation advancement.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

As AI Like It: Neural Network Models for Recognizing Website Aesthetics

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Ivan Arkanov,
  • Maxim Bakaev

摘要

Test-driven development has obtained nearly universal adoption in back-end programming, and now front-end testing automation sees extensive progress. To advance beyond auto-checking the purely “syntactic” correction of code and designs of user interfaces (UIs), it is necessary to predict quantifiable dimensions of user satisfaction and experience (UX) – particularly, visual aesthetics. In this task, neural networks (NNs) already largely overpower the approaches based on cognitive architectures that were popular in the past. However, datasets in Human-Computer Interaction (HCI) generally have limited sizes, while pre-trained models for predicting aesthetic impressions do not transfer well between different website domains and user groups. In our paper, we employ openly available datasets by Calista (overall 42,000 aesthetic evaluations for about 500 website screenshots) and compare available NN models (AlexNet, CaffeNet, and NIMA-MobileNet) with varying numbers of layers and the representation of the output. Fine-tuning AlexNet with another open dataset, Web-AVA, yielded the best improvement of RMSE by about 5%. The results of our study might be of interest for both AI/ML researchers and practitioners, and for web designers who seek to catch up with the today’s NN-based testing automation advancement.