基于低秩矩阵学习的多源排序融合关键技术研究

Ranking of retrieved/recommended results is a fundamental research topic in information retrieval and recommendation systems. Ranking plays an important role in many real world applications such as search engines, product recommendation, accurate advertisement. Rank aggregation is a task that firstly obtains multiple individual rank lists via multi-source features or different base rankers, and then combines these rank lists to get a better one. Since rank aggregation is unsupervised and doesn't need manually labeled data, it can be applied to various applications such as meta-search, recommendation systems, machine translation, and prediction fusion in machine learning tasks. There are three main challenges for rank aggregation methods: the input individual rank lists are incomplete, noisy and with large number of items. In this proposal, we focus on a novel rank aggregation method via low-rank matrix learning. We propose to use a low-rank matrix to model the comparison matrix of the items to be ranked, and use a structured-sparse matrix to model the errors in the base rankers. Then we investigate the corresponding objective function and optimization algorithm for the new rank aggregation model. To tackle the issue of large-scale input data in real world scenarios, we propose a divide-and-conquer approximation solution for scalable rank aggregation. Finally we analyze the recovery conditions for rank aggregation and the theoretical bounds on the approximation errors of the divide-and-conquer scheme.

在信息检索和推荐系统中，对检索/推荐结果的排序是最核心的研究问题。排序是搜索引擎，产品推荐、精准广告投放等重要应用中的关键技术。排序融合是指分别通过多种不同来源的特征或多种算法获得搜索/推荐结果的多个基本排序，然后把这些基本排序融合得到一个更好的排序。排序融合是一种不需要人工标注数据的无监督方法，因而可适用于互联网元搜索，各种推荐系统，机器翻译，机器学习中的预测结果融合等问题中。实际应用中输入数据不完整，高噪声和规模大给已有的排序融合方法带来了重要挑战。本课题拟研究基于低秩矩阵学习的排序融合方法。首先，拟提出用低秩矩阵来建模待排序对象间的先后序关系，并用结构化稀疏的矩阵来建模对象关系中的噪声结构；针对基于低秩矩阵的排序融合模型和噪声结构，提出相应的目标函数和优化算法；针对实际应用中数据规模大的问题，拟提出基于分治思想的排序融合的近似学习算法；分析排序融合的成功条件和分治近似错误。

在信息检索和推荐系统中，对检索/推荐结果的排序是最核心的研究问题。排序融合方法广泛应用于互联网元搜索，推荐系统，机器学习中的预测结果融合等问题中。实际应用中输入数据不完整，高噪声和规模大给已有的排序融合方法带来了重要挑战。..本课题研究基于低秩矩阵学习的排序融合方法，并将其扩展到多视图聚类、多任务学习等结构相似的机器学习问题上。首先，提出用低秩矩阵来建模待排序对象间的先后序关系（或聚类对象的相似/不相似关系），并用结构化稀疏的矩阵来建模对象关系中的噪声结构；针对基于低秩矩阵的排序融合模型（或聚类模型）和噪声结构，提出相应的目标函数和优化算法；针对实际应用中数据规模大的问题，提出基于分治思想的近似学习算法。另外，本课题还研究了基于深度学习的哈希方法及其在检索和排序中的应用。..本课题的代表性成果包括：在AAAI，CVPR，ECCV等权威国际会议上发表论文5篇（其中CCF A类会议论文3篇）；在IEEE Transactions期刊上发表（含录用）论文5篇。其中，AAAI 2014的论文“Supervised Hashing for Image Retrieval via Image Representation Learning”在Google Scholar上的引用次数为195次，CVPR 2015的论文“Simultaneous Feature Learning and Hash Coding with Deep Neural Networks”在Google Scholar上的引用次数为183次；AAAI 2014的论文“Robust Multi-View Clustering via Low-rank and Sparse Decomposition”在Google Scholar上引用次数为78次。