引入无监督表示学习和堆叠双向GRU的预测模型

Health Prediction Model with Unsupervised Feature Learning and Stacked Bidirectional GRU

可穿戴设备正侵入我们的生活,通过其内嵌的传感器采集大量数据为获取人体的各项指标提供参考,但相关数据的敏感性和隐私性局限了可获取的疾病标签,为健康预测增添了一定的难度.为了解决可获得的人体活动数据多但是疾病标签匮乏的问题,提出了一种称为act2vec的无监督表示学习模型,该方法从原始活动数据中学习时序数据的特征表示,通过分布式表示来挖掘可用于疾病预测任务的活动模式.考虑到数据的特征矩阵无法全面反映样本特性的问题,通过考虑活动等级的周期性,嵌入代表活动级别的序数关系,利用噪声对比估计构建表示学习损失函数,提取数据特征矩阵.最后,引入所学习的特征矩阵构建堆叠双向GRU模型以进行疾病预测.2个数据集上的实验表明了所提方法的有效性.

Wearable devices are invading our lives,and a large amount of data is collected through its embedded sensors to provide a reference for obtaining various indicators of the human body.However,the sensitivity and privacy of related data limit the available disease labels,which adds a certain degree of difficulty to health prediction.In order to solve the problem of much available human activity data but the lack of disease tags,an unsupervised representation learning model called act2vec is proposed.The model learns the feature representations of time series from the original activity data and mines activity pattern which can be used for disease prediction tasks through distributed representations.Taking into account the problem that the feature matrix of the data cannot fully reflect the characteristics of the samples,we consider the periodicity of the activity level and embed the ordinal relationship representing the activity level.The representation learning loss function is constructed by utilizing noise contrast estimation and the data feature matrix can be extracted.Finally,the learned feature matrix is introduced to construct a stacked bidirectional GRU model for the disease prediction.Experiments on 2 datasets show the effectiveness of the proposed method.