Research On Hybrid Recommendation Algorithm Based On Deep Learning

In addition to connecting people all over the world,the Internet has profoundly changed the way people live and work.With more and more information resources are available to people through the Internet,the problem of information overload is becoming more and more serious.Recommendation system is one of the most effective ways to solve this problem.It learns user preferences by collecting explicit and implicit feedback from users on items and quickly retrieves interesting content for them.The recommendation algorithm is the core module of the recommendation system,which determines the performance of the entire system.Hybrid recommendation algorithm is a commonly used recommendation algorithm at present.The most important principle of this kind of algorithm is to combine multiple types of recommendation algorithms to avoid the weaknesses of their respective algorithms,to make the recommendation results more accurate.There are also some problems of common hybrid recommendation algorithm to be solved.For example,although the hybrid recommendation algorithm using deep neural networks can learn high-order feature combination information,it is difficult to deal with dense numerical features directly.Although the hybrid recommendation algorithm using the gradient boosting decision tree can handle dense numerical features well,and has certain interpretability and feature selection capabilities,but it is difficult to handle massive data and online updates.In view of the above problems,the work of this thesis is summarized as follows:(1)This thesis proposes a hybrid recommendation model LFDNN,which consists of four modules.The Light GBM module applies gradient boosting decision trees to the feature processing part of the model,improving the model's ability to handle dense numerical features.In the shallow model module,the FM model is introduced to explicitly model finite-order feature crosses,which strengthens the expressive ability of the model.The deep neural network module uses a fully-connected feedforward neural network to allow the model to obtain more high-order feature crosses information and mine more data patterns in the features.The Fusion module allows the shallow model and the deep model to obtain a better fusion effect.The model comparison experiment,parameter influence experiment and ablation experiment are carried out on two real data sets respectively to verify the validity of the proposed model.(2)This thesis also proposes another hybrid recommendation model,TKDDNN,which is divided into two modules in parallel.The GBDT module uses gradient boosting decision trees to process dense numerical features,allowing the model to learn more information from the dense numerical features.At the same time,combined with knowledge distillation technology,the tree obtained in the gradient boosting decision tree is converted into a neural network output to adapt to online updates and large-scale data.The LFDNN2 module uses an independent hybrid recommendation model as a sub-part of TKDDNN.The shallow part uses a domain-aware factorization machine to further enhance the ability of feature crosses,and the deep part uses a deep neural network to deal with sparse category features.The model comparison experiments are carried out on two advertising datasets to verify the validity of the model.