【作者】 高强；

【导师】 郭菲； 陈列伟；

【作者基本信息】 天津大学 ， 计算机技术， 2019， 硕士

【摘要】 稀疏数据是在数据集中绝大多数数值缺失或者为零的数据,如何挖掘稀疏数据特征之间的隐含关系从而对其进行预测分析是本文的主要研究问题。目前,主流的挖掘特征间隐含关系的方法无法较全面地抽取特征间隐含关系进行更深层次的学习。通过在特征学习过程中对特征和特征之间的隐含关系进行丰富和扩充,从而提高对特征间隐含关系的学习能力。本文提出一种基于特征之间隐含关系的稀疏预测方法,结合丰富的特征信息参与特征间隐含关系的自动学习。首先,本文提出基于特征之间隐含关系的稀疏预测模型框架FIRM,将多路特征信息融入特征之间隐含关系的学习中。其次,本文针对特征的可传递性和可重用性,提出基于短链接和层损失的InteractionNN模型,该模型利用特征的可传递性构建特征之间的低阶特征交叉,同时利用特征的可重用性自动学习特征之间的高阶特征交叉。最后,针对特征之间低级别的特征间隐含关系,本文提出基于混合低阶特征交叉提取的MINN模型,该模型在学习低阶特征交叉时引入三阶特征交叉,丰富低阶特征交叉的多样性。本文对提出的模型在标准数据集上进行大量的对比实验,实验结果表明本文提出的模型可以较好地挖掘特征之间的隐含关系。InteractionNN模型和MINN模型均可自动学习较为充分的特征间隐含关系,并且取得较好的性能表现。本文提出的方法可满足实际应用层面上的需求,可较好地自动学习特征之间的隐含关系,同时可避免人力的浪费。更多还原

【Abstract】 Sparse data is the data with most values missing or zero in the dataset,how to mine the implicit relationship between sparse data features and predictive analysis of sparse data is the main research issue of this paper.At present,the mainstream methods of mining implicit relationship between features cannot extract the implicit relationship between features for deeper learning.By enriching and expanding the implicit relationship between features in the feature learning process,the learning ability of the implicit relationship between features is improved.This paper proposes a sparse prediction method based on the implicit relationship between features,and combines rich feature information to participate in the automatic learning of implicit relationships between features.Firstly,this paper proposes a sparse prediction model framework FIRM based on the implicit relationship between features,which integrates multi-channel feature information into the learning of implicit relationships between features.Secondly,aiming at the transitibility and reusability of features,this paper proposes an InteractionNN model based on shortcut connection and layer loss.This model uses the transitivity of features to construct low-order feature interactions between features,and learns high-order feature interactions between features by utilizing the reusability of features.Finally,for the low-level implicit relationship between features,this paper proposes a MINN model based on hybrid low-order feature cross extraction,which introduces third-order feature interactions when learning low-order feature interactions,enriching the diversity of low-order features interaction.In this paper,a large number of comparative experiments are carried out on the proposed dataset.The experimental results show that the proposed model can better mine the implicit relationship between features.InteractionNN model and MINN model can automatically learn the more implicit relationships between features and achieve better performance.The method proposed in this paper can meet the requirements at the practical application level,and can automatically learn the implicit relationship between features automatically,while avoiding the waste of manpower.更多还原