摘要：

推荐算法是一种用于解决信息过载问题的方法, 引文推荐通过引文上下文能够自动匹配候选论文列表。现有基于神经引文网络模型在引文上下文数据预处理的过程中, 存在文本噪声和上下文学习不充分的问题。为此, 提出一种基于层间融合滤波器和社交神经引文网络的推荐算法FS-Rec。首先, 利用具有层间融合滤波器的BERT模型预处理引文上下文, 在频域内从所有频率中提取有意义的特征, 缓解引文上下文数据的噪声, 同时在频域中对多层信息进行融合, 增强上下文表示学习的能力; 然后, 在引文作者嵌入中引入社交关系, 与其他引文信息嵌入通过编码器获得表示, 将这些表示与经过BERT预训练的引文上下文表示进行融合, 得到最终表示; 最后, 根据最终表示生成引文文本预测。实验结果表明, 相较于现有的上下文引文推荐模型, FS-Rec在2个基准数据集arXivCS和PubMed取得了更高的召回率和平均倒数排名(MMR), 证明了模型的有效性。

关键词: 滤波器, 自注意力机制, 基于Transformer的双向编码器表示, 引文推荐, 预训练语言模型

Abstract:

Recommendation algorithms address information overload by automatically matching citation recommendations with a list of candidate papers, utilizing the citation context. Existing neural citation network models encounter issues, such as text noise and insufficient context learning, while preprocessing citation context data. Therefore, this study proposes a recommendation algorithm based on interlayer fusion filter and social neural citation network FS-Rec. First, the citation context is preprocessed using a Bidirectional Encoder Representations from Transformers (BERT) model equipped with interlayer fusion filters. This process involves extracting meaningful features from all frequencies within the frequency domain, thereby mitigating noise in citation context data. Multilayer information is simultaneously fused within the frequency domain, thus enhancing the capability of contextual representation learning. Social relationships are then introduced into citation author embeddings, and representations are obtained through an encoder along with other citation information embeddings. These representations are subsequently fused with the pre-trained citation context representations using BERT to obtain the final representation. Finally, citation text predictions are generated based on the ultimate representation. Experimental results demonstrate that compared with existing context-based citation recommendation models, FS-Rec achieves higher recall rates and Average Reciprocal Ranks(MRR) on two benchmark datasets, arXivCS and PubMed, thereby proving the effectiveness of the model.

Key words: filter, self-attention mechanism, bidirectional encoder representation from Transformer, citation recommendation, pre-trained language model