Deep Learning-based Heterogeneous Temporal Events Prediction And Its Application

Many fields such as intelligent transportation and health care produce massive timeseries data,which is heterogeneous and dynamic.Using the methods such as deep learning to effectively mine the complex time dependencies in time-series data to predict heterogeneous temporal events and provide decision support for application services in related fields has become a research hotspot.Firstly,existing works mainly focus on solving the heterogeneous problems due to the multi-variate and multi-rate of time-series data while ignoring the domain knowledge’s guiding role in event prediction.Secondly,artificial intelligence algorithms such as deep learning inevitably lack interpretability,knowing "results" without presenting "causes." An excellent prediction model still faces the challenge to its process rationality and is not easy to be recognized by experts.This paper proposes a heterogeneous temporal events prediction algorithm that combines domain knowledge and model interpretability to solve the above problems.Firstly,it integrates specific domain knowledge and employs deep neural networks to mine the complex nonlinear dependencies in heterogeneous time-series data,thereby improving heterogeneous temporal events’ prediction ability.Then an interpretability module is introduced to interpret the prediction process in the temporal events prediction task.The proposed algorithm is verified and evaluated in the ICU patient endpoint prediction task.The thesis contents are as follows.Firstly,analyze the characteristics of heterogeneous time-series data and perform preprocessing operations such as data completion and data annotation on the clinical medicine domain’s datasets.Secondly,integrate the prior knowledge in clinical medicine to construct a weight module,and design a P-BiLSTM algorithm based on bidirectional Long Short-Term Memory model.The fusion of clinical domain knowledge improves deep neural networks’ learning ability,so P-BiLSTM has better accuracy than six baseline methods in predicting ICU patient endpoints on two real-world ICU datasets(PhysioNet and MIMIC Ⅲ).Thirdly,we take advantage of the SHAP model constructed with information gain ratio to improve the P-BiLSTM algorithm and design an interpretable heterogeneous temporal events prediction algorithm(PI-BiLSTM).The experimental results show that in the ICU patient endpoint prediction task,the feature importance ranking is consistent with the prior clinical knowledge,verifying the PI-BiLSTM algorithm’s interpretability.