基于多任务学习的应用商店客户识别模型研究

客户识别模型因为其广泛的应用背景、迫切的现实需求，正成为客户关系管理领域的热点研究问题。应用商店客户识别本质上是一个高维类别不平衡的“多产品联合客户识别”问题，然而现有客户识别模型主要是针对单个产品的客户识别问题设计的，忽略了识别任务之间的相关性。围绕“多产品客户联合识别”这一核心问题，本项目首先将多任务学习技术引入到客户识别中，提出了多任务客户联合识别模型框架；进而通过类别不平衡处理策略和特征选择策略研究，将GMDH神经网络和集成学习技术结合构建了一种能够处理类别不平衡并能在高维度数据情况下高效稳定地进行特征选择的多任务集成神经网络；最后将模型应用到解决实际的应用商店客户识别问题中去。研究成果将传统的单产品客户识别模型扩展为多产品联合客户识别模型，是对客户识别模型的理论方法创新，同时所构建的多任务集成神经网络模型为应用商店客户关系管理提供了有用的工具。

Customer identification is becoming a hot issue in the field of customer relationship management because of its wide application background and urgent practical need. Customer identification in applications store is essentially a high-dimensional imbalanced “multiple product joint customer identification” problem. However, existing customer identification is mainly designed for the single product scenario, which ignores the relatedness between learning tasks. Around the core issue of “multiple product joint customer identification”, the project first introduces the technique of multi-task learning into customer identification and builds the model framework of multi-task customer identification. Then, after studying class imbalance and feature selection strategies, we combine GMDH neural network and ensemble learning techniques together to build a multi-task model that could deal with class imbalance and select the feature efficiently and stably under high dimension. Finally, the model is applied to solve real problems in the application stores. This project extends the traditional single product customer identification model to multi-product joint customer identification model, which is the model innovation. Meanwhile the proposed multi-task ensemble neural network model provides a useful tool for application stores...