基于图注意力网络的下一个购物篮推荐

GRAPH ATTENTION NETWORK FOR NEXT-BASKET RECOMMENDATION

购物篮是用户在一次购物中一系列物品的集合。在下一个购物篮推荐中,可以根据用户的历史购物篮序列推荐出用户接下来购买的物品集合。受图注意力网络的启发,利用该网络建模购物篮中的物品联系,得到含有物品关系的购物篮表示;为了建模用户的短期兴趣,设计一个固定大小的移动窗口,融合最近几个连续购物篮的用户偏好;融合经过循环神经网络的篮子表示、全局和短期的用户购物偏好得到最终的推荐。在两个公开数据集上进行实验,结果表明该算法可有效提高下一个购物篮推荐精度。

A basket is a collection of items that a user buys at a time.In the next basket recommendation,we recommend the next set of items the user will buy based on their historical basket sequence.Inspired by the graph attention network,we used this network to model the relation of items in the basket,and the expression of the basket containing the relation of items was obtained.In order to model the short-term interests of users,a fixed-size mobile window was designed,which integrated the user preferences of several successive shopping baskets.The basket representation through the recurrent neural network,global and short-term user basket preference were fused to get the final recommendation.A large number of experiments on two public datasets demonstrate that this algorithm can effectively improve the accuracy of next-basket recommendation.