基于多源异质信息融合的推荐系统研究

With the development of Internet, there are more and more available heterogeneous historical behavior data related to users and items. It provides an opportunity to alleviate the data sparsity and cold start problem of recommender systems, and boost the performance in recommendation accuracy as well. However, research on heterogeneous information fusion based recommender systems is still in its infancy and is far from being solved mainly for three reasons: 1) the given meta-patterns suffer from the poor expression power in semantics; 2)the heterogeneous information is inefficiently represented; 3)preference feature aggregation lacks effectivity. In this project, we explore how to automatically discover accurate meta-patterns, how to represent heterogeneous information networks based on network embedding, and how to aggregate preference features via the factorization machine. The project aims to lay the theoretical foundation and develop key techniques for heterogeneous information fusion based recommender systems. Specifically, it is expected to propose an automatic meta-pattern recognition algorithm, design a novel method to represent heterogeneous information networks under the guidance of meta-patterns, and develop a feature aggregation strategy based on the factorization machine. The research of our project is doomed to inaugurate novel insights for the field of heterogeneous information fusion based recommender systems, which exerts a profound and far-reaching influence and is of great practical application value.

随着互联网的发展，推荐系统中有关用户、项目的历史行为数据变得更具多源性和异质性，为缓解推荐系统中的数据稀疏性和冷启动问题、提高推荐准确性提供了机遇。然而，融合多源异质信息的推荐系统研究尚处于起步阶段，仍然存在给定元语义模式语义表示能力有限、多源异质信息表示效率较低、偏好特征融合缺乏有效性等问题。本项目拟通过探索元语义模式发现方法，自动挖掘语义表示能力更为精准的元语义模式；研究基于网络嵌入技术的异质信息网络表示方法，提高多源异质信息的处理效率；探究基于因子分解机模型来融合特征，提高融合多源异质信息推荐的有效性。本项目的总体目标是建立基于多源异质信息融合的推荐系统的若干理论基础和关键技术，提出元语义模式自动发现方法，提出元语义模式指导下的异质信息网络表示方法，提出基于因子分解机的偏好特征融合策略。为基于多源异质信息融合的推荐系统研究提供一种新思路，研究工作具有重要的科学意义和实际应用价值。

针对当前基于异质信息网络的推荐系统中存在的关键问题：（1）给定元语义模式语义表示能力有限；（2）多源异质信息表示效率较低；（3）偏好特征融合缺乏有效性。本项目借助深度学习、图挖掘等技术，实现自动挖掘语义表示能力更为丰富、精准的元语义模式，提高多源异质信息的处理效率，构建多源偏好特征融合策略。具体地，本项目通过探索元语义模式发现方法，自动挖掘语义表示能力更为精准的元语义模式；研究了基于网络嵌入技术的异质信息网络表示方法，提高多源异质信息的处理效率；探究了基于因子分解机模型来融合特征，提高了融合多源异质信息推荐的有效性。基于真实数据集的对比实验结果验证了本项目所提融合多源异质信息的推荐系统在提高推荐准确度、提高运算效率等方面的有效性。为推动基于多源异质信息融合的推荐系统走向实际应用提供理论支撑和技术支持，具有重要的研究意义和广阔的应用背景。