CAJ | 학술논문

基于正则化奇异值分解(RSVD)的推荐算法在预测准确性上具有明显优势，但存在计算量大的缺点，一定程度上限制了其在实际工程中的应用。针对这个问题，该文提出一种基于标签迁移学习的改进RSVD推荐算法，在相对稠密的辅助数据集合上根据标签信息来提取用户/项目特征，然后将用户/项目特征应用到RSVD算法中，在目标数据集合上进行评分预测。在 MovieLens 数据集合上的实验结果表明，该算法能够缓解数据稀疏性问题，显著降低预测误差(约0.01 RMSE)，同时减少约50%的训练时间。
기우정칙화기이치분해(RSVD)적추천산법재예측준학성상구유명현우세，단존재계산량대적결점，일정정도상한제료기재실제공정중적응용。침대저개문제，해문제출일충기우표첨천이학습적개진RSVD추천산법，재상대주밀적보조수거집합상근거표첨신식래제취용호/항목특정，연후장용호/항목특정응용도RSVD산법중，재목표수거집합상진행평분예측。재 MovieLens 수거집합상적실험결과표명，해산법능구완해수거희소성문제，현저강저예측오차(약0.01 RMSE)，동시감소약50%적훈련시간。
The recommender algorithm based on Regularized Singular Value Decomposition (RSVD) has significant advantages in predictive accuracy, while it is computationally intensive, which limits greatly its application to engineering projects. To address this issue, an improved algorithm based on tag transfer learning is proposed. It leverages tag information in the relatively denser auxiliary dataset to extract user/item features, which are further used in the RSVD approach in order to make recommendation in the target dataset. Experiments on MovieLens datasets show that the proposed algorithm can handle the sparsity issue effectively, achieve far better prediction results (reducing about 0.01 RMSE), and save about 50%training time at the same time.