Abstract
Live streaming platforms have become a dominant form of online contentconsumption, offering dynamically evolving content, real-time interactions, andhighly engaging user experiences. These unique characteristics introduce newchallenges that differentiate live streaming recommendation from traditionalrecommendation settings and have garnered increasing attention from industry inrecent years. However, research progress in academia has been hindered by thelack of publicly available datasets that accurately reflect the dynamic natureof live streaming environments. To address this gap, we introduce KuaiLive, thefirst real-time, interactive dataset collected from Kuaishou, a leading livestreaming platform in China with over 400 million daily active users. Thedataset records the interaction logs of 23,772 users and 452,621 streamers overa 21-day period. Compared to existing datasets, KuaiLive offers severaladvantages: it includes precise live room start and end timestamps, multipletypes of real-time user interactions (click, comment, like, gift), and richside information features for both users and streamers. These features enablemore realistic simulation of dynamic candidate items and better modeling ofuser and streamer behaviors. We conduct a thorough analysis of KuaiLive frommultiple perspectives and evaluate several representative recommendationmethods on it, establishing a strong benchmark for future research. KuaiLivecan support a wide range of tasks in the live streaming domain, such as top-Krecommendation, click-through rate prediction, watch time prediction, and giftprice prediction. Moreover, its fine-grained behavioral data also enablesresearch on multi-behavior modeling, multi-task learning, and fairness-awarerecommendation. The dataset and related resources are publicly available athttps://imgkkk574.github.io/KuaiLive.