计算机科学与探索
計算機科學與探索
계산궤과학여탐색
JOURNAL OF FRONTIERS OF COMPUTER SCIENCE & TECHNOLOGY
2014年
6期
751-759
,共9页
推荐系统%协同过滤%大数据%稀疏矩阵%联合聚类%迁移学习
推薦繫統%協同過濾%大數據%稀疏矩陣%聯閤聚類%遷移學習
추천계통%협동과려%대수거%희소구진%연합취류%천이학습
recommender system%collaborative filtering%big data%sparse matrix%co-clustering%transfer learning
针对传统协同过滤推荐(collaborative filtering recommendation,CFR)受数据聚类预处理,评分矩阵稀疏性影响较大和多个评分矩阵之间不能知识迁移的问题,提出了一种基于联合聚类和评分矩阵共享的协同过滤推荐方法,以提高推荐系统精度和泛化能力。该方法首先通过联合聚类对原始评分矩阵进行用户和项目两个维度的聚类;然后对评分矩阵进行分解并取得共享组级评分矩阵;最后利用共享组级评分矩阵和迁移学习方法进行评分预测。对MovieLents和Book-Crossing两个数据集进行了仿真实验,结果表明该方法相比传统方法平均绝对误差减少近8%,有效地提高了协同过滤推荐的预测精度,为协同过滤推荐的应用提供借鉴。
針對傳統協同過濾推薦(collaborative filtering recommendation,CFR)受數據聚類預處理,評分矩陣稀疏性影響較大和多箇評分矩陣之間不能知識遷移的問題,提齣瞭一種基于聯閤聚類和評分矩陣共享的協同過濾推薦方法,以提高推薦繫統精度和汎化能力。該方法首先通過聯閤聚類對原始評分矩陣進行用戶和項目兩箇維度的聚類;然後對評分矩陣進行分解併取得共享組級評分矩陣;最後利用共享組級評分矩陣和遷移學習方法進行評分預測。對MovieLents和Book-Crossing兩箇數據集進行瞭倣真實驗,結果錶明該方法相比傳統方法平均絕對誤差減少近8%,有效地提高瞭協同過濾推薦的預測精度,為協同過濾推薦的應用提供藉鑒。
침대전통협동과려추천(collaborative filtering recommendation,CFR)수수거취류예처리,평분구진희소성영향교대화다개평분구진지간불능지식천이적문제,제출료일충기우연합취류화평분구진공향적협동과려추천방법,이제고추천계통정도화범화능력。해방법수선통과연합취류대원시평분구진진행용호화항목량개유도적취류;연후대평분구진진행분해병취득공향조급평분구진;최후이용공향조급평분구진화천이학습방법진행평분예측。대MovieLents화Book-Crossing량개수거집진행료방진실험,결과표명해방법상비전통방법평균절대오차감소근8%,유효지제고료협동과려추천적예측정도,위협동과려추천적응용제공차감。
In view that the traditional collaborative filtering recommendation (CFR) is affected largely by the data clustering preprocess and the sparsity of rating-matrix, cannot allow knowledge-sharing across multiple rating matrices, this paper puts forward a new collaborative filtering method of combining the co-clustering with rating-matrix sharing to improve the forecasting accuracy and generalization ability. Firstly, it uses the co-clustering method to divide the raw rating-matrix into clusters by two dimensions of users and items. Secondly, it factorizes the rating-matrix and obtains a cluster-level rating-matrix. At last, it predicts the rating-matrix using the cluster-level rating-matrix and transfer learning method. A simulation experiment using databases of MovieLents and Book-Crossing is carried out. The results show that this method reduces nearly 8%for the average absolute error value compared to the traditional CFR, and improves the prediction accuracy of recommender system. This method provides references for the collaborative filtering recommendation.