Research On Personalized Sleep Prediction Based On Group-model

In recent years,smart phones have become more powerful and can carry more functions.But at the same time,the power consumption of smart phones is also increasing.The development of battery technology is relatively slow,and Moore's Law is not followed,resulting in a slow increase in the supply capacity of mobile phone batteries,which cannot meet the needs of mobile phone energy consumption and limits the battery's life.Therefore,under the background that battery technology cannot make breakthroughs,it is of great significance to study how to effectively reduce the power consumption of mobile phones.When users sleep at night,they don't actually use the phone,but there are still many unnecessary processes and modules running in the background of the phone,such as background applications and GPS positioning modules.These background unnecessary processes or modules will occupy massive transmission energy consumption and calculation energy consumption.If the mobile phone terminal can accurately identify whether the user is in a sleep state,then when the user is in a sleep state,unnecessary processes or modules can be terminated in the background,and the mobile phone enters a low-power mode to improve battery life.The user's behavior and state can be obtained from the rich sensor data and human-computer interaction information on the mobile phone,so the information can be used to infer whether the user is in a sleep state.When the process or module is terminated in the background,user experience needs to be considered,and there can be no more accidental killing operations.When using commonly indicators such as accuracy and recall,it is not possible to avoid frequent jumps in prediction results,and frequent jumps in prediction results will bring more accidental killings,which will seriously affect the user experience.Therefore,the wake-up rate and coverage rate indicators are designed.The wake-up rate focuses on the power saving efficiency of the mobile phone,and the coverage rate focuses on the user experience.Existing sleep prediction is mainly in the health field,usually predicting the length of sleep,sleep quality and other quantities related to user health,often using models such as decision trees,SVM,LSTM networks,etc.In this topic,using these models will cause a lot accidental killing operations and cannot achieve the effect of good user experience.According to the above problems,this paper proposes a model based on the evolutionary LSTM neural network to construct a group model and an individual model respectively.The group model is oriented to group data,as a factory model and a back-off model,and considers the characteristics of most users when designing.The individual model is built based on the group model,focusing on personalization and dynamics on the basis of the group model.Using the LSTM neural network can better process the characteristics of sequence data,and use the differential evolution algorithm for training to directly optimize the comprehensive target of wake-up rate and coverage.The coverage rate represents the proportion of time the user is in low-power mode when the user sleeps.Coverage rate means high power-saving efficiency,wake-up rate pays attention to the number of transitions of prediction results,and low wake-up rate means good user experience.In the group model,compared with the traditional classification model,the coverage rate is increased by about 5%when the wake-up rate is low.For individual models,a strategy based on clustering and fine-tuning is adopted to reflect personalization,and concepts drift detection and sliding window training are used to monitor the dynamic changes of user data.The experimental results show that the above method can better reflect the personalized and dynamic changes of individual user data.