| Online shopping has rapidly grown with the continuous improvement and development of online payment and modern logistics.In the era of traffic supremacy,commodity recommendation algorithms provide personalized recommendations to better meet user needs and preferences.More and more researchers are developing various high-quality recommendation algorithms for commodity,but there are still some shortcomings.Firstly,the problems of data sparsity and cold-start have not been well solved.Negative sampling in the data pre-processing stage of machine learning models plays a crucial role in handling incomplete datasets,the current negative sampling methods still suffer from redundant exploration operations and high computational complexity,resulting in low sampling efficiency and low effectiveness of negative samples,which cannot obtain high-quality negative items.Secondly,previous recommendation algorithms only combine user long-term and short-term preferences in a simple and static way,without further exploring specific contextual feature information,which leads to inefficient user modeling.This thesis proposed a personalized recommendation algorithm based on improved negative sampling and dynamic purchase intention to address these issues,including the following two parts:Negative sampling method had redundant exploration operations and high computational complexity,resulting in low sampling efficiency and negative sample effectiveness,and cannot obtain high-quality negative samples.In response to this problem,an improved knowledge graph negative sampling method based on a dual attention mechanism(IKGDA)was proposed.Firstly,the high-quality representations of KG nodes and user-item nodes were preprocessed through the node embedding representation module.Since rich knowledge graph structure was introduced,it was helpful to improve the execution efficiency of exploration operations.Then,through the dual attention mechanism,the KG entity related to the current item node exposed to the user was estimated using the knowledge graph attention,and the item was selected as the least interesting negative example for the user from its neighbors using the item attention,thereby reducing redundant path exploration and reducing exploration complexity.Finally,the parameter of IKGDA was optimized using an optimization function to generate a high-quality negative sample paired with a userpositive interaction item.The proposed method was applied to the MF matrix factorization recommendation algorithm,and compared with mainstream negative sampling methods such as MCNS and KGPL on three public datasets.The experimental results shown that the proposed method achieved significant improvements in the RECALL@20 and NDCG@20 metrics.Commodity recommendation algorithm based on dynamic purchase intention(DPI-Rec)was proposed to address the low modeling efficiency issue,which was caused by static combination of long-term and short-term user preferences.Firstly,in order to deal with the time interval between different purchase behaviors in the user behavior sequence to solve the problem of time irregularity,this thesis proposed to use the improved gating mechanism in LSTM to convert the time interval to the future stage;then,in order to deal with the item interval between different purchase behaviors in the user behavior sequence to solve the problem of irregular semantic content,this thesis combined the predictive embedding of the target item and used the attention mechanism to dynamically filter irrelevant actions in the user behavior sequence.The recommendation performance of DPI-Rec algorithm was effectively improved by joint training of time interval,project interval and predictive embedding of target items(user dynamic purchase intention).Finally,the user ’s short-term intention and long-term preference in specific situations were integrated through the adaptive fusion module.Comparative experiments with mainstream recommendation algorithms,such as CL4 SRec and DIEN,on three public datasets demonstrated significant improvements in F1 and AUC metrics for DPI-Rec algorithm. |
