【作者】 李刚；

【导师】 王光；

【作者基本信息】 辽宁工程技术大学 ， 电子信息硕士（专业学位）， 2023， 硕士

【摘要】 随着数据处理技术的深入研究与发展,如今推荐系统已在工业界得到了广泛的应用,并在多种领域充分体现了其商业价值,但同时其在许多预测场景中仍然存在数据稀疏性与冷启动问题。以往的推荐工作大多是监督学习,这不仅增加了数据标注成本及时间成本,且导致算法的适应能力较弱,面对一个全新的应用场景难以快速构造一个高效的推荐系统。为此,在借鉴相关研究成果的基础上,提出了一种结合对比学习的自监督推荐算法(Self-supervised Recommendation Algorithm Combined with Constrastive Learning,CLSRec)。算法通过构造完形填空预测与子序列预测两个基于正负例的对比学习任务来完成自监督学习。首先,对于一个特定场景下的用户交互序列,按照对应任务要求分别随机遮掩掉其个别项目与子序列,将其依次通过嵌入层与自注意力层来预测其缺失项,预测到的项目即视为预测结果;其次,针对用户序列被遮掩的项目,自动构造相应预测结果的正负例,结合预测结果与对比学习损失函数完成梯度学习,学习到正确的项目语义特征与空间向量编码;最后,在下游任务中利用上游训练好的项目编码来微调模型进行个性化推荐,获得满足用户偏好的推荐结果。通过与多个算法进行对比实验,CLSRec分别在HR、NDCG、MRR三个指标上获得了更好的表现,平均获得了16.2%的HR指标与15.3%的NDCG指标以及12.0%的MRR指标提升,并通过消融实验分析及参数可视化分析过程,验证了算法的合理性与有效性。该论文有图20幅,表7个,参考文献63篇。更多还原

【Abstract】 With the in-depth research and development of data processing technology,the recommendation system has been widely used in the industry,and fully embodies its commercial value in a variety of fields,but at the same time,it still has data sparsity and cold start problems in many prediction scenarios.Most of the previous recommendation work is supervised learning,which not only increases the cost of data annotation and time,but also leads to the weak adaptability of the algorithm,which makes it difficult to construct an efficient recommendation system quickly in the face of a new application scenario.To do this,this paper proposes a Self-supervised Recommendation Algorithm Combined with Constrastive Learning(CLSRec)on the basis of relevant research achievements.The algorithm accomplishes self-supervised learning by constructing two contrast learning tasks based on positive and negative examples,cloze prediction and subsequence prediction.First,for the user interaction sequence in a specific scene,individual items and sub-sequences are randomly covered respectively according to the corresponding task requirements,and the missing items are predicted through the embedding layer and the self-attention layer in turn.The predicted items are regarded as the predicted results.Secondly,the positive and negative examples of the corresponding prediction results are automatically constructed for the projects covered by the user sequence,and the gradient learning is completed combined with the prediction results and the contrast learning loss function,so that the correct project semantic features and spatial vector coding are learned;Finally,in the downstream task,the project code trained in the upstream is used to fine-tune the model for personalized recommendation and obtain the recommendation results that meet user preferences.Compared with multiple algorithms,CLSRec has achieved better performance in HR,NDCG and MRR,respectively,with an average increase of 16.2% in HR,15.3% in NDCG and 12.0% in MRR.Through the ablation experiment analysis and parameter visualization analysis process,The rationality and validity of the algorithm are verified.The paper has 20 pictures,7 tables,and 63 references.更多还原