Abstract: Guild survival has a positive effect on improving the game user’s activity and retention rate. Currently the main approach to the problem is based on a two-class classification process, which fails to make full use of the longitudinal data reflecting the state change and the survival trend of a guild. A joint model for longitudinal and survival data is adopted to predict guild survival state. The model fully utilizes the features of the attributions and member behaviors for a guild. The experiments show that the performance of the joint model for the longitudinal and survival data improves the overall performances of 56.6% than Cox proportional hazards model, and the model is more precise than that of the commonly used classification algorithms, e. g. the overall performance of this approach is 11. 9% better than that of logistic regression. Moreover, the following conclusions can be drawn from the experiment. The first is that a more hierarchical guild architecture leads to a more stable guild, since the standard deviation of the right levels of the members has a positive impact on the guild survival. The second is that, the diversity among the behavior of guild members results in longer guild survival time, according to the effects of the standard deviation of private chat times and the PK times. The third is that, the survival time for guild shows a negative effect, and this means the longer the guild is living, the less conducive to the survival of the guild.

Key words: classification algorithms, game guild, joint model, survival prediction

摘要： 行会生存对提高游戏用户的活跃度和留存率有着积极的作用。目前行会生存分析方法是使用分类法，即把行会是否生存看作一个二分类问题来处理，其未能充分利用行会纵向数据，不能及时反映行会的状态变化和生存趋势。采用纵向-生存联合模型，充分利用游戏行会纵向的状态变化特征和成员行为特征，预测行会的生存状态。实验表明，纵向-生存联合模型相比传统的Cox比例风险模型，综合性能提高了56.6%，相比分类算法提高了预测性能，如逻辑回归提高了11.9%。实验中发现：成员权利等级标准差对行会生存呈现了正向的影响，说明了行会内成员权利等级有着良好的分布对行会的生存有着重要作用；成员私聊次数标准差和成员PK次数标准差对行会生存有着积极的影响，说明行会成员行为差异性的重要性；生存时间对行会有着负向的影响，即行业已生存时间越长，越不利于行会的生存。

关键词: 分类算法, 游戏行会, 联合模型, 生存预测