Note: clustering tech in Recommendation Systems

未分类

• [1] Can Movies and Books Collaborate ? Cross-Domain Collaborative Filtering for Sparsity Reduction
  
  • cross domain collaborate. 研究了不同领域之间评分的协同转换，用成熟领域的评分来预测新领域。在本文中使用的是电影领域和书籍领域
  • establish a bridge between the two rating matrices and transfer useful rating patterns from the movie rating matrix to the book rat- ing matrix.
• [2] A Scalable Collaborative Filtering Framework based on Co-clustering
  
  • 对用户物品同时聚类，通过类簇评分均值，加入个体偏移量后的协同过滤框架。
  • 硬聚类
  • 快，实时，适用于新加入用户和物品的场合
  • Empirical comparison of our approach with SVD, NNMF and correlation-based collaborative filtering techniques indicates comparable accuracy at a much lower computational effort.
  • 描述NMF算法：The matrix factorization approaches includeSVD [13] and NNMF-based [17, 10] filtering techniques that pre- dict the unknown ratings based on a low rank approxi- mation of the original ratings matrix.
  • 描述矩阵分解算法的优点：Unlike correlation-based methods, the matrix factorization techniques treat the users and items symmetrically and hence, handle item synonymy and sparsity in a better fashion.
  • 很多co-clustering的前作，减少稀疏性和维度：There has been a lot of work on co-clustering [8, 7], most of which was focused on handling sparsity and dimensionality reduction issues in text and micro-array analysis
  • the effects of changes in the ratings ma- trix are localized which makes it possible to have efficient incremental updates、
  • The predicted ratings can then be used to generate the top-N recommendations if required. A natural approach to this rating prediction problem is to assume that the matrix has a certain low parameter structure, deduce the parameters of the structure based on the available ratings so that a certain loss function is min- imized, and then use a matrix reconstruction based on this structure for predicting the missing values. To
• [3] Predictive Discrete Latent Factor Models for Large Scale Dyadic Data
  
  • simultaneously incorporates the effect of covariates and estimates local structure that is induced by interactions among the dyads through a discrete latent factor model
  • tried simultaneously hard and soft co-cluster method
• [4] cHawk : An Efficient Biclustering Algorithm based on Bipartite Graph Crossing Minimization
  
  • 理论上分析了谱聚类和二分图交叉最小化的联系。provide connection between spectral partioning and cross minimization
• [5] A Framework for Simultaneous Co-clustering and Learning from Complex Data
  
  • 提出了一个基于模型的co-clustering算法并在多领域，包括推荐领域进行了应用的测试
  • 硬聚类
  • This simultaneous approach is better than independently clustering the data first and then building classification models.
• [6] A Generalized Maximum Entropy Approach to Bregman Co-clustering and Matrix Approximation
  
  • 提出了一种同时考虑类簇内部均值和行与列向量全局均值的联合聚类方法，并在推荐系统上进行了应用
• [7] Employing User Attribute and Item Attribute to Enhance the Collaborative Filtering Recommendation
  
  • 协同（硬）聚类，在使用user-item关联矩阵的基础上使用了用户和物品的属性
  • The personalization of clustering technique recommendation is not so good as other CF. In some cases the clusters are less accurate than the memory- based algorithms
• [8] A COLLABORATIVE FILTERING APPROACH COMBINING CLUSTERING AND NAVIGATIONAL BASED CORRELATIONS Clustering based Recommender
  
  • 提出了PSN-CF（Pam clustering on Similarities and Navigational based-CF），只聚类了用户，聚类方式通过UI矩阵得到UU相似度矩阵，在UU相似度上使用PAM（Partitioning Around Medoids）进行聚类
• [9] PAC-Bayesian Analysis of Co-clustering and Beyond
• [10] A Clustering Approach for Collaborative Filtering Recommendation Using Social Network Analysis
  
  • 社交网络分析，置信度分析
• [11] Beyond "local", "categories" and "friends": clustering foursquare users with latent "topics"
  
  • 根据用户历史轨迹对用户进行聚类，（LDA话题模型）得到高于地点，categories等信息的聚类
  • Introduction和相关工作中提到
  • 通过历史轨迹预测社交网络朋友关系[1-23,1-20]
  • 城市特征/现象[1-24]，neighborhood现象[1-7,8]
  • 得出了结论：人类流动特征，人大部分的地理轨迹一般会集中在很小的地理区域中[1-16~18,6] [直接引用文中的结论可以么[1]]
  • Yet while heterogeneity in a population can to a large extent be explained by closeness of social and geodesic distances [1-4]
  • Similarly, user displacement, or distance be- tween two successive check ins, follows a power law which
• [12] Improve Collaborative Filtering Through Bordered Block Diagonal Form Matrices
  
  • 协同block clusteing。(Approximate) Bordered Block diagonal form structure, 使用二分图和矩阵两种表现形式。定义了分块矩阵形式。这种形式兼顾了热门用户/物品对每个block的影响。
• [13] Localized Matrix Factorization for Recommendation based on Matrix Block Diagonal Forms
  
  • block clustering
  • 12和13是13年同一批人做的
• [14] An evolutionary clustering algorithm based on temporal features for dynamic recommender systems
• [15] Biclustering neighborhood-based collaborative filtering method for top- n recommender systems
• [16] Improving Co-Cluster Quality with Application to Product Recommendations
• [17] Hierarchical Co-Clustering : Off-line and Incremental Approaches
• [18] A Hybrid Multigroup Coclustering Recommendation Framework
• [19] WEMAREC : Accurate and Scalable Recommendation through Weighted and Ensemble Matrix Approximation Categories and Subject Descriptors
• [20] Local Item-Item Models For Top-N Recommendation
  
  • 对用户进行单独聚类。
  • 改变Label，loss function 同时考虑了局部信息、全局矩阵和用户关系
• [21] Collaborative Filtering Bandits
• [22] An Architecture for Privacy Preserving Collaborative Filtering on Web Portals
  
  • 提出了一个协同过滤推荐框架用以解决协同过滤问题中的用户隐私问题，其中应用了交叉最小化的协同聚类的聚类方式。
• [23] 两阶段联合聚类协同过滤算法
  
  • soft block clustering
  • 当行和列同时具有相关性时,应当考虑使用联合聚 类,因为无论从哪一个维度进行单独聚类时,都会忽略另一维的相关信息