Abstract: Mobile phone game providers gain profits by selling virtual props in games. This paper describes an event in game log data of players as an instance and represents the states of players’ prop purchases by labels, so that the game props recommendation is modeled as multi-instance multi-label learning (MIML) problem. On the basis of this abstraction, an MIMLfast algorithm is used to recommend mobile phone game props and a semi-supervised learning part is integrated to improve the performance of recommendation. The results of experiments conducted on offline data sets and a real mobile phone game show that the MIML-based game props recommendation brings a remarkable increase of game profits.

Key words: machine learning, multi-instance multi-label learning (MIML), semi-supervised learning, recommendation

摘要： 手机游戏提供商通过在游戏中销售虚拟道具来获得收益。将游戏玩家日志数据中每个事件描述为一个示例，玩家对多种游戏道具的购买状态表示为多个标记，从而将游戏道具推荐问题抽象为多示例多标记学习问题。在此基础上，将快速多示例多标记学习算法用于手机网络游戏道具推荐，并利用半监督学习提升推荐性能。离线数据集以及实际在线手机网络游戏实验结果表明，基于多示例多标记学习的游戏道具推荐技术带来了游戏营收的显著增长。

关键词: 机器学习, 多示例多标记学习（MIML）, 半监督学习, 推荐