[1]LU Xiangzhi,SUN Fuzhen,WANG Shaoqing,et al.Long sequence recommendation algorithm based on user memory matrix[J].CAAI Transactions on Intelligent Systems,2023,18(3):517-524.[doi:10.11992/tis.202110003]
Copy
CAAI Transactions on Intelligent Systems[ISSN 1673-4785/CN 23-1538/TP] Volume:
18
Number of periods:
2023 3
Page number:
517-524
Column:
学术论文—智能系统
Public date:
2023-07-05
- Title:
-
Long sequence recommendation algorithm based on user memory matrix
- Author(s):
-
LU Xiangzhi; SUN Fuzhen; WANG Shaoqing; DONG Jiawei; WU Xiangshuai
-
College of Computer Science and Technology, Shandong University of Technology, Zibo 255000, China
-
- Keywords:
-
memory network; hierarchy; long-term interest; short-term interest; long short-term memory network; gated recurrent unit; long sequence recommendation; session recommendations
- CLC:
-
TP309.2
- DOI:
-
10.11992/tis.202110003
- Abstract:
-
Traditional recurrent neural networks, such as long-term and short-term memory networks (LSTM) and gated recurrent unit (GRU), have limited memory capacity and inflexible access to memory data, which have inherent shortcomings in capturing features of longer sequences. The memory network has the characteristics of storing long-term memory, and the access to memory data is more flexible and changeable. Therefore, this paper introduces the memory network in the session-based recommendation algorithm. In this paper, we design a hierarchical recommendation model, which is divided into two layers. The first layer is the session-level GRU model, which is used to characterize the sequence of current session and predict the next item. The second layer is the user-level memory network model, which is used to describe the changes in users’ long-term interests. The model proposed in this paper can effectively capture the short-term and long-term interests of users and thus improve the performance of recommendations. The experiments on public data sets demonstrate that the proposed hierarchical network with user memory (HNUM) algorithm has 4% increase in both recall rate and mean inverse ranking improvement relative to hierarchical gated recurrent unit (HGRU) for a performance improvement comparison of 10 sessions versus 5 sessions.