基于深度神经网络的搜索引擎点击模型构建

发布时间：2018-03-24 15:51

本文选题：异质化结果　切入点：深度神经网络　出处：《中文信息学报》2017年05期

【摘要】：随着富媒体展现形式被越来越多地引入搜索交互界面,搜索引擎的结果页面呈异质化和二维模块展现形式,这对传统的点击预测模型提出了巨大的挑战。针对这一情况,我们对实际搜索引擎结果页面的多模态结果进行了分析,构建了一个结合深度神经网络和点击模型的框架,该框架既包含了神经网络的特性,又利用了点击模型的预测能力。我们希望利用这个框架挖掘出多模态信息与文本信息之间的相关性,使之具有描述异质化结果和二维模块展示形式的能力。实验表明,我们的框架相较于传统的点击模型在点击预测性能上有显著提升,但由于搜索引擎的多模态结果内容复杂,仅利用多模态结果的底层特征,即使使用深度神经网络,从中能够挖据出的语义相关性较弱。
[Abstract]:As rich media presentation is increasingly introduced into the search interactive interface, the search engine results pages are heterogeneous and two-dimensional module presentation, which poses a huge challenge to the traditional click-prediction model. We analyze the multi-modal results of the actual search engine results page, and construct a framework combining the depth neural network and the click model, which includes the characteristics of the neural network. We also use the prediction ability of click-through model. We hope to mine the correlation between multi-modal information and text information by using this framework, so that it has the ability to describe the heterogeneous results and display the form of two-dimensional module. The experimental results show that, Compared with the traditional click-through model, our framework has a significant improvement in the performance of click-prediction, but because of the complexity of the multi-modal results of search engines, only the underlying features of the multi-modal results are utilized, even though the depth neural network is used. The semantic correlation that can be extracted from it is weak.
【作者单位】：智能技术与系统国家重点实验室清华信息科学与技术国家实验室(筹)清华大学计算机系;
【基金】：国家自然科学基金(61622208,61532011,61472206) 国家973计划(2015CB358700)
【分类号】：TP391.3

【相似文献】