Anime Recommendation System Based On Deep Learning

Anime,as a type of film presented in animated form,has always been beloved by people.With the rapid development of the internet today,a massive number of anime works are released on various online anime platforms every day.These anime websites often recommend the same content to all users,making it difficult for users to quickly find anime works that interest them from the vast pool of anime resources.This paper designs and implements a deep learning-based anime recommendation system that can provide personalized recommendations to users and help them quickly discover anime that they are interested in.The deep recommendation model used in this system has strong expressive and fitting abilities,which can achieve good recommendation results.The system consists of two stages:recall and ranking,with a focus on the deep recommendation model used.In the recall stage,a graph embedding-based recall algorithm using random walk is employed.This algorithm utilizes random walk on the item relationship graph to obtain multiple sequences of anime,and these sequences are then fed into the Item2vec model for training to generate anime embeddings.The LSH algorithm is used to find similar anime,and experimental results show that this recall algorithm has a higher F1 score,indicating better recommendation performance.In the ranking stage,an MLP model with embedded features is used.This algorithm improves the input layer by embedding sparse categorical features and concatenating them with other features as input for model training.Comparative experiments demonstrate that this ranking model achieves higher AUC,MRR,NDCG@5,and NDCG@10 scores compared to other deep learning ranking models.Vue and Flask are selected as the frontend and backend frameworks for the anime recommendation system.The system architecture includes the presentation layer,data layer,and strategy layer.The functional modules of the system mainly include the feature engineering module,offline computation module,online recommendation module,model evaluation module,and interface display module.The real-time recommendation module meets the real-time requirements of the recommendation system.Based on users’ online click-through rates,it utilizes the Flink streaming processing platform to promptly incorporate the collected user behavior information into the subsequent recall layer for recommendation.The system uses a distributed message queue,Kafka,for storing and accessing user behavior data,and utilizes Spark to calculate and obtain the input features required by the recommendation model.Finally,a user-friendly anime recommendation system is implemented,and after comprehensive system testing,the system functions properly and performs well.