Abstract

Determining the prevalence of alcohol-related content on social media is important to guide education initiatives and interventions in this space. We aimed to assess the performance of the pre-developed alcoholic beverage identification deep learning algorithm (ABDILA) to automatically quantify alcoholic beverage prevalence in user-generated Instagram images. 6,121 images were gathered from Instagram using ‘Splendour in the Grass’ related hashtags, an Australian music festival. These images were manually annotated as containing beer, champagne, wine, or anything else. The images were subsequently run through ABIDLA, which made predictions on their same categorical contents. We then assessed overall model accuracy (relative to human annotations), model accuracy of alcohol-containing images (overall accuracy and across beverage categories), and visually inspected images to extract common features of congruent- or mis-categorisations. While overall accuracy was high, congruent classifications were heavily skewed towards non-alcohol images. The algorithm consistently overestimated the number of images containing alcoholic beverages, and inspection revealed that these false positives were largely driven by image context and colour. While such algorithms show early promise as a rough automated estimation tools for large datasets on social media, this study highlights some critical improvements and directions for applying pre-trained algorithms in this space.

摘要

确定社交媒体上酒精相关内容的流行程度对于指导该领域的教育计划和干预措施非常重要。我们旨在评估预先开发的酒精饮料识别深度学习算法 (ABDILA) 的性能，以自动量化用户生成的 Instagram 图像中的酒精饮料流行率。使用澳大利亚音乐节“草丛中的辉煌”相关标签从 Instagram 收集了 6,121 张图片。这些图像被手动注释为包含啤酒、香槟、葡萄酒或其他任何东西。这些图像随后通过 ABIDLA 运行，后者对其相同的分类内容进行了预测。然后，我们评估了整体模型准确性（相对于人类注释）、含酒精图像的模型准确性（整体准确性和跨饮料类别），和视觉检查图像以提取一致或错误分类的共同特征。虽然整体准确度很高，但一致的分类严重偏向于非酒精图像。该算法始终高估了包含酒精饮料的图像数量，并且检查显示这些误报很大程度上是由图像上下文和颜色驱动的。虽然此类算法显示出作为社交媒体上大型数据集的粗略自动化估计工具的早期前景，但本研究强调了在该领域应用预训练算法的一些关键改进和方向。该算法始终高估了包含酒精饮料的图像数量，并且检查显示这些误报很大程度上是由图像上下文和颜色驱动的。虽然此类算法显示出作为社交媒体上大型数据集的粗略自动化估计工具的早期前景，但本研究强调了在该领域应用预训练算法的一些关键改进和方向。该算法始终高估了包含酒精饮料的图像数量，并且检查显示这些误报很大程度上是由图像上下文和颜色驱动的。虽然此类算法显示出作为社交媒体上大型数据集的粗略自动化估计工具的早期前景，但本研究强调了在该领域应用预训练算法的一些关键改进和方向。