Abstract:

Aiming at the problems of sparse data, low accuracy and single recommendation in the application of traditional collaborative filtering technology in network recommendation system, this paper proposes a deep neural network recommendation model named DSM combining skip-gram item embedding model and weighted loss function. Firstly, the proposed model uses three layers of ReLU to regress the output vectors, which improves the recommendation accuracy without using additional information. Secondly, this paper uses the Skip-gram model for item embedding to obtain a denser representation vector, which reduces the computation. Moreover, the weighted loss function is used to train the parameters of the deep neural network, which balances the popularity of the recommended items and ensures the novelty. Finally, the experimental results in the APP dataset and the Last.fm dataset show that the proposed model has a certain improvement in accuracy and diversity compared to existing methods when recommending applications and songs.

Key words: recommendation system, data sparse, Skip-gram, weighted loss function, deep neural network

摘要：

针对网络推荐系统中传统的协同过滤技术在实际应用中存在数据稀疏、导致准确率低、推荐单一性等问题，提出一种结合Skip-gram项目嵌入和加权损失函数的深度神经网络的推荐模型DSM。采用了3层ReLU层对输出向量进行回归，在未使用附加信息的前提下提高了推荐精度；利用Skip-gram进行项目嵌入得到更稠密的表示向量，减少了计算量；并且使用加权损失函数训练深度神经网络的参数，平衡了推荐项目的受欢迎程度，保证了新颖性。在APP数据集和Last.fm数据集的实验结果表明，DSM模型在推荐应用程序和歌曲时，准确性和多样性方面相比现有方法均有一定的提高。

关键词: 推荐系统, 数据稀疏, Skip-gram, 加权损失函数, 深度神经网络