面向智能环境的活动模式迁移学习

针对智能环境中活动模式的学习和挖掘花销大、难以实际操作等问题,提出了能够有效地将已有活动模式迁移到新环境的整体框架。迁移学习框架将活动模式的迁移过程分解为轨迹的迁移和触发持续时间的迁移,首先对已有活动模式中的活动轨迹以及触发持续时间模糊化;然后采用备选轨迹生成(ATSG)算法在新环境中生成备选轨迹集;最后采用相似度计算(SC)算法进行活动模式中的轨迹与备选轨迹间的匹配,利用活动轨迹映射(TM)算法和触发持续时间迁移(TDT)算法对活动信息进行迁移,从而在新环境中得到活动模式。理论分析和实验结果表明,相比于基于频繁模式挖掘得到活动模式的方法,本文方法大幅度地降低了得到活动模式所需的时间开销,同时,利用本文方法获取的活动模式取得了较好的活动识别效果。

Transfer learning for activity pattern in smart environment

Due to huge costs and operational difficulties in learning and mining activity patterns in smart environment, we propose a framework for activity pattern transfer. The process of activity pattern transfer is divided into two parts: trajectory transfer and trigger duration transfer. Firstly, fuzzify the trajectories and trigger durations in the existing activity patterns. Secondly, generate alternative trajectory set through alternative path set generation (ATSG) algorithm in a new environment. Finally, map the trajectories from activity patterns and trajectories from alternative trajectory set through similarity computing (SC) algorithm, and transfer activities' information through trajectory mapping (TM) algorithm and trigger duration transfer (TDT) algorithm. Thus, activity patterns are transferred to the new environment. Theoretical analysis and experimental results show that, compared to frequent pattern mining, the method in our paper can substantially reduce time overhead of obtaining activity patterns in new environment. At the same time, it performs well in recognizing activities.