高级检索

    基于多级注意力机制网络的app流行度预测

    app Popularity Prediction with Multi-Level Attention Networks

    • 摘要: 移动app流行度预测从运营策略优化到短期广告投资都具有重要意义.利用应用市场提供的丰富数据,挖掘不同特征指标与流行度之间的动态关联,从而预测未来一段时间内app的流行度变化过程及其原因,对于开发者、投资商和应用市场三方都具有应用价值.然而,app流行度高度动态变化,其影响因素十分复杂,包括app自身迭代演化、用户反馈、同类产品的市场竞争等.目前,针对app流行度建模与预测的研究工作相对较少,大多通过构造人工特征并构建与流行度的关联来进行预测,在模型的计算性能、预测精度和结果可解释性等方面存在可提升的空间.因此针对app流行度建模与预测提出一种基于注意力机制的深度神经网络模型DeePOP,并针对复杂影响因素进行分级建模.通过时间级自序列模块捕获对历史流行度的长期依赖,利用局部、全局特征层级模块构建影响特征与流行度的非线性关系.同时,注意力机制为不同模块提供自适应能力,以捕获与流行度变化最相关的历史状态并针对预测结果提供一定的解释.实验结果表明:与现有流行度预测方法相比,DeePOP能够快速高效地进行app流行度建模与预测,预测均方根误差为0.089.

       

      Abstract: The popularity prediction of mobile apps provides substantial value to a broad range of applications, ranging from operational strategy optimization to targeted advertising and investment. This work includes leveraging the rich data provided by the app market to mine the dynamic correlation between different factors and popularity, so as to predict the app popularity over the next period of time, which creates great value for developers, investors and the app market. However, the evolution of app popularity is highly dynamic, and its influence factors are very complex, including the iterative evolution of the app itself, user feedback, and competition for similar products and so on. At present, there are relatively few research studies on app popularity modeling and prediction. Most of them construct artificial features and capture its association with popularity, and there is room for improvement in terms of computational performance, prediction accuracy, and interpretability of results. In this paper, we propose DeePOP, an attention based neural network for app popularity modeling and prediction, which performs hierarchical modeling for complex influence factors. First, we propose the time-level self-sequence module to capture the long-term dependence on historical popularity, and propose the local and global feature level modules to capture the nonlinear relationship between features and app popularity. Second, the attention mechanisms provide adaptive capabilities for different modules to capture most relevant historical states and provide explanation for prediction. Last, the experimental results show that DeePOP outperforms the state-of-the-art methods and the root mean square error of prediction reaches up to 0.089.

       

    /

    返回文章
    返回