Recommendation algorithms are a method to address the issue of information overload. Citation recommendation automatically matches a candidate list of papers through the context of citations. Existing citation recommendation models suffer from low sensitivity to the temporal aspect of citation context representation learning and the presence of noise in complex text data. To address these issues, we propose a global citation recommendation algorithm named CTAF-GCR, based on CTA-LSTM and a filtering enhancement learning module. Firstly, the algorithm utilizes a CTA-LSTM layer to process citation context data, adaptively assigning weights to each point in the sequence through a temporal attention mechanism to better capture the temporality of citations, thus enhancing the model’s temporal sensitivity. Meanwhile, the filtering enhancement learning module suppresses noise by extracting meaningful features from all frequencies within the frequency domain, thereby alleviating noise in citation context data, and then obtaining the final representation. Subsequently, the final representation is used for citation text prediction. Compared to existing citation recommendation models, experimental results on three benchmark datasets, ACL-200, ACL-600, and RefSeer, demonstrate the model’s effectiveness, achieving higher recall rates and mean reciprocal ranks.