基于异构信息网络的分类算法推荐方法研究

项目名称： 基于异构信息网络的分类算法推荐方法研究

项目编号： No.61502378

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 其他

项目作者： 王广涛

作者单位： 西安交通大学

项目金额： 22万元

中文摘要： 在实际应用中，对于给定分类问题，面对众多的候选分类算法，如何为用户推荐合适的算法来解决该分类问题就显得尤为重要，同时也是数据挖掘领域挑战性问题之一。已有的推荐模型通常基于元学习来构建，利用元数据来描述分类数据集特征和算法性能间的关系。这种解决方式忽略了数据集间的关系以及算法间的关系，且难以将用户需求纳入到推荐模型构建过程当中。项目采用异构信息网络来对算法和数据集间的关系进行建模，充分考虑数据集间和算法间的关系，利用异构信息网络分析技术来研究算法推荐问题进而构建分类算法推荐模型。项目的研究内容包括：算法－数据集异构信息网络构建，信息网络随分类算法发展的进化演变，推荐模型构建以及用户需求指导下的分类算法推荐。项目主要贡献在于突破传统基于元学习推荐模型的局限，较早运用异构信息网络及其分析技术实现分类算法推荐模型的构建，有利于研究人员更好地理解算法处理能力和算法在实际应用中的有效运用。

中文关键词： 分类算法；异构信息网络；元模式；算法推荐

英文摘要： In practical applications, for a given classification problem and many candidate classification algorithms, recommending appropriate algorithms for the problem is very important and also one of the challenging problems in the field of data mining. The existing recommendation models are usually constructed based on meta-learning, which employs meta-data to model the interaction between the characteristics of classification data set and the performance of the classification algorithms. The meta-learning based methods ignore the interaction between the data sets and the interaction between algorithms. And it is also difficult to construct the recommendation models while considering the user requirements. In this project, we employ the heterogeneous information network to model the interaction between data sets and classification algorithms. Then, the techniques used for heterogeneous information network are used for classification algorithm recommendation model construction. The research contents include: algorithm-data heterogeneous information network construction, network evolution with the development of the classification technique, algorithm recommendation model construction and the user-guided classification algorithm recommendation. The main contribution of the project it that, the project utilizes the analysis technologies over the heterogeneous information network for classification algorithm recommendation earlier, and it breaks through the limitations of traditional meta-learning based methods. This is beneficial to both the researchers to understanding the processing capacity of classification algorithms and users to applying classification algorithm more effectively.

英文关键词： Classification algorithm;Heterogeneous information network;Meta-schema;Algorithm recommendation