A recommendation algorithm based on fullink consistency optimization

Abstract Recommendation systems have evolved into multi-purpose and modular developments. Due to the complexity of the recommendation pipeline, this inevitably leads to inconsistent goals among the modules and the overall pipeline. To improve the consistency of the recommendation systems pipeline, we propose a knowledge-based learning framework for full pipeline consistency (COKA). This is the first work to combine knowledge graphs and apply them to the full recommendation pipeline.First, in order to enhance the precision of the ranking model with the highest accuracy in the recommendation pipeline, we introduce the user's historical comments and use the Named Entity Recognition (NER) system to identify and combine entity linking (EL). The system maps the identified entities in the first step to the corresponding entities in Wikipedia.Next, we construct a subgraph that depends on the extracted entities and related entities. Here we use GCN[1] to represent the contextual information of the subgraph. On this basis, we introduce the infonce loss[2] of contrastive learning to help the model further identify the metric expression of the sample. At the same time, in constructing the sample stream of contrastive learning, we introduce a global extra queue in addition to the batch simple[3] to further help alleviate the SSB[4] problem.Finally, we learn the final ranking of the precision ranking through a small dual-tower and bring the expression of the precision ranking model into the entire pipeline through a simple structure to enhance the overall consistency of the recommendation pipeline.