基于服装图像视觉特征的冷启动问题缓解

冷启动问题是协同过滤推荐算法中被广泛关注的问题,它的存在严重影响协同过滤算法的推荐质量. 提出深度卷积神经网络提取的服装商品图像视觉特征用于计算用户对新商品喜好度的方法来缓解冷启动问题,并利用矩阵分解模型估算用户对服装商品的评分. 通过从服装商品图像视觉特征到商品特征向量的映射函数计算新商品的特征向量,给出了两种映射函数形式:K最近邻映射和线性映射. 实验结果表明,服装图像视觉特征能够有效缓解协同过滤算法冷启动问题.

Mitigation of Cold-Start Problem Based on Visual Features of Clothing Images

The cold-start problem is a classic problem which has widely been concerned in the collaborative filtering recommendation algorithm. The problem seriously affects the recommendation quality of the collaborative filtering algorithm. This paper proposes a way to alleviate the cold-start problem by using the visual feature of the clothing product image learned by the deep convolutional neural network. The paper uses the matrix factorization model to estimate the users’ score on clothing items. In this paper,the items feature vector is calculated by a mapping function from the clothing product image visual feature to the items feature vector. The paper mentions two forms of mapping functions:K nearest neighbor mapping and linear mapping. The experimental results show that the visual feature of clothing image can effectively alleviate the cold-start problem of collaborative filtering algorithm.