基于用户长短期偏好的序列推荐模型

Sequential Recommendation Model Based on User’s Long and Short Term Preference

针对现有序列推荐模型忽略了不同用户的个性化行为,导致模型不能充分捕获用户动态偏好而产生的兴趣漂移等问题,提出了一种基于用户长短期偏好的序列推荐模型(Sequential Recommendation Model Based on User’s Long and Short Term Preference,ULSP-SRM)。首先,根据用户的序列中交互物品的类别和时间信息生成用户的动态类别嵌入,进而有效建立物品之间的关联性,并且降低数据的稀疏性;其次,根据用户当前点击物品和最后一项点击的时间间隔信息生成个性化时序位置嵌入矩阵,模拟用户的个性化聚集现象,以更好地反映用户偏好的动态变化;然后,将融合了个性化时序位置嵌入矩阵的用户长期偏好序列以会话为单位输入门控循环单元中,生成用户的长期偏好表示,并通过注意力机制将用户长短期偏好进行融合,生成用户的最终偏好表示,从而达到充分捕获用户偏好的目的;最后,将用户最终偏好表示输入推荐预测层进行下一项推荐预测。在Amazon公开数据集的7个子集上进行实验,采用AUC(Area Under Curve)值、召回率和精确率指标进行综合评估,实验结果表明,所提模型的表现优于其他先进基准模型,有效地提升了推荐性能。

Aiming at the problem that the existing sequence recommendation model ignores the personalized behavior of different users,the model cannot fully capture the interest drift caused by users’dynamic preferences,a sequence recommendation model based on users’long and short term preferences(ULSP-SRM)is proposed.Firstly,the dynamic category embedding of the user is generated according to the category and time information of the interactive items in the user’s sequence,thereby effectively establishing the correlation between the items and reducing the sparsity of the data.Secondly,according to the time interval information of the user’s current clicked item and the last clicked item,a personalized time series position embedding matrix is generated to simulate the user’s personalized aggregation phenomenon and better reflect the dynamic change of user preference.Then,the user’s long-term preference sequence fused with the personalized time-series position embedding matrix is input into the gated recurrent unit in units of sessions to generate the user’s long-term preference representation,and the user’s long and short term preferences are fused through the attention mechanism to generate the final preference representation of the user,to achieve the purpose of fully capturing the user’s preference.Finally,the final user preference representation is input to the recommendation prediction layer for the next recommendation prediction.Experiments are carried out on seven subsets of Amazon public data set,and the area under curve(AUC),recall rate and precision rate indicators are used for comprehensive evaluation.Experimental results show that the proposed model outperforms other advanced benchmark models,effectively improving recommended perfor-mance.