An efficient and resource-aware hashtag recommendation using deep neural networks

The goal of this research is to design a system that can predict and recommend hashtags to users when new images are uploaded. The proposed hashtag recommendation system is called HAZEL (HAshtag ZEro-shot Learning). Selecting right hashtags can increase exposure and attract more fans on a social media platform. With the help of the state-of-the-art deep learning technologies such as Convolutional Neural Network (CNN), the recognition accuracy has improved significantly. However, hashtag prediction is still an open problem due to the large amount of media contents and hashtag categories. Using single machine learning method will not be sufficient. To address this issue, we combine image classification and semantic embedding models to achieve the expansion of recommended hashtags. In this research, we show that not all hashtags are equally meaningful, and some are not suitable in recommendation. In addition, by periodically updating semantic embedding model, we ensure that the hashtags being recommended follow the latest trends. Since the recommended hashtags have not received any training examples in the first place, it fulfills the concept of Zero-shot learning. We demonstrate that our system HAZEL can successfully recommend hashtags that are the most relevant to each image input by applying our design to a larger scale of image-hashtag pairs on Instagram.