TAGAN:一种融合细粒度语义特征的学术论文对抗推荐算法

TAGAN:an academic paper adversarial recommendation algorithm incorporating fine-grained semantic features

学术论文推荐旨在为用户提供个性化的论文资源,针对协同过滤方法面临数据高度稀疏和缺乏负样本的问题,提出了一种融合细粒度语义特征的学术论文对抗推荐模型--TAGAN(title and abstract GAN)。首先,基于具有语义特征的标题和摘要,使用卷积神经网络(CNN)提取标题的全局特征,并构建一个双层的长短期记忆(LSTM)网络分别对摘要的单词序列和语句序列建模,同时,引入注意力机制将标题和摘要进行语义上的关联。然后,将论文的语义特征融入基于生成对抗网络(GAN)的推荐框架中并进行训练,其生成模型会拟合用户的兴趣偏好,能有效替代负采样过程。最后,通过在公开数据集上的实验对比,TAGAN在各个指标上都优于基线模型,验证了TAGAN的有效性。

Academic paper recommendation aims to provide users with personalized paper resources.Collaborative filtering methods face the problems of highly sparse data and lack of negative samples.Considering the above chal-lenges,an academic paper recommendation algorithm TAGAN(title and abstract GAN)which incorporated fine-grained semantic features was presented.Firstly,based on titles and abstracts provide abundant semantic features,convolutional neural networks(CNN)was used to extract the global features of the titles,a two-layer long and short-term memory network(LSTM)was built to model abstract words separately.At the same time,the attention mechanism was proposed to associate the title and the abstract semantically.Then,the semantic features of the paper were integrated into the recommendation framework based on generative adversarial network(GAN).The generative model will fit the user’s interest preferences and can effectively replace the negative sampling process.Finally,through the experimental comparison on the public dataset,TAGAN is better than the baseline models in all indica-tors,which verifies the effectiveness of TAGAN.