随着移动终端设备（包括智能手机、可穿戴设备等）的智能化与集成化不断增强，充分利用其内嵌的各类传感器，如麦克风、陀螺仪、WiFi、蓝牙等，可以实现对用户行为的全面感知与透彻理解。在此基础上，构建用户精准画像对于重点人群监测、精准商业推荐、个性化服务等各类应用均具有十分重要的作用和意义。然而在现实中，用户行为特性/规律往往分散在真实的物理世界、虚拟的信息空间以及隐秘的内心世界中，因此，要实现用户的精准画像必须把散落在不同空间和维度的多源感知数据有效融合起来，构建统一的、融合的用户画像框架。基于此，本文开展了基于多源感知数据的用户精准画像研究，具体研究内容包括以下几个方面：

（1）人格心理空间：通过手机感知的多源数据对大五人格模型和心理模型进行刻画。利用方差选择法和皮尔逊相关系数法提取与人格心理问卷结果相关性显著的特征，使用递归特征消除法进行特征选择，用训练后的逻辑回归模型、支持向量机模型和随机森林模型分别对大五人格和心理状态进行预测，得到九维向量表示大五人格模型（5维）和心理模型（4维）的预测结果。

（2）信息空间：使用手机产生的通讯数据对用户社会关系进行刻画。通讯数据包含手机用户唯一识别标志，基于用户间通信频次和时长构建周内周末的社交网络，研究用户在通讯网络中社交信任度和影响力的表现，成为衡量用户关系紧密度的重要指标。

（3）物理空间：移动场景下产生的位置感知数据对用户移动分布关系、位置语义和移动行为特征进行刻画。分析不同用户的蓝牙连接分布，可发现用户的移动分布关系；利用TF-IDF分类方法提取位置语义，对用户访问位置偏好加以标记；对获取的蓝牙连接频次和时长建立周内周末的社会关系网络，区别用户周内周末的移动行为特征。

在此基础之上，为了有效验证本文所构建的基于多源数据的用户精准画像方法，将之应用到移动群智感知应用服务中。具体而言，将用户的移动行为和社会关系画像分别应用于任务离线分发和在线分发中，离线任务分发阶段提出了基于效用函数的任务分配方法，随着任务数量增多，相比于传统的任务分发方法任务完成率和任务分发成功率平均提高了10%；在线任务分发阶段提出了基于社会关系画像的任务转发方法，包含社会关系、用户信誉值和物理距离三个要素的任务转移实验中，比仅包含用户信誉值和物理距离要素的任务转移实验中任务完成率平均提高了15%。

With the increasing intelligentization and integration of mobile terminal devices (including smart phones, wearable devices, etc.), the comprehensive perception and thorough understanding of user behaviors can be realized by making full use of all kinds of embedded sensors, such as microphone, gyroscope, WiFi, Bluetooth, etc.On this basis, the construction of accurate user portraits has a very important role and significance for key population monitoring, accurate business recommendation, personalized service and other applications.In reality, however, the user behavior characteristics/law is often scattered in the real physical world, virtual information space, and secret inner world, therefore, in order to realize accurate portrait of users must be found scattered in different space and dimension of multi-source data fusion effectively perception, the construction of a unified, integration of user picture frame.Based on this, this paper carries out a research on accurate user portrait based on multi-source perception data. Specific research contents include the following aspects:

(1) Personality psychological space: The big five personality model and psychological model are depicted through multi-source data of mobile phone perception.Choice method of variance, and Pearson correlation coefficient method is used to extract the characteristics of the significant correlation with personality psychological questionnaire results, using recursive feature elimination method for feature selection, use after training logistic regression model, support vector machine (SVM) model and random forest model respectively to forecast the big five personality and psychological state, get nine dimensional vector said big five personality model (5 dims) and the predictions of a mental model (4 dims).

(2) Information space: Use the communication data generated by mobile phones to depict users' social relations.The communication data contains the unique identification marks of mobile phone users. Based on the frequency and duration of communication between users, social networks are constructed on weekends in a week. The performance of social trust and influence of users in communication networks is studied, which becomes an important indicator to measure the closeness of user relationships.

(3) Physical space: The location perception data generated in the mobile scene describes the user's movement distribution relationship, location semantics and movement behavior characteristics.The distribution of bluetooth connection of different users can be analyzed to find the mobile distribution relation of users.Tf-idf classification method is used to extract location semantics and mark the user's access location preference.Establish a social relation network on the weekend of a week for the frequency and duration of acquired Bluetooth connection to distinguish the mobile behavior characteristics of users on the weekend of a week.

On this basis, in order to effectively verify the multi-source data-based user accurate portrait method constructed in this paper, it is applied to mobile swarm intelligence application service.Specifically, the user's mobile behavior and social relations portrait respectively applied to distribute task offline and online distribution, offline tasks distribution stage task allocation method based on utility function are presented, with the task, task completion compared with the traditional tasks distribution method and task distribution success rate increased by 10% on average;In the online task distribution stage, a task forwarding method based on social relationship portrait was proposed. In the task transfer experiment involving social relationship, user's credit value and physical distance, the task completion rate in the task transfer experiment involving only user's credit value and physical distance was increased by 15% on average.