Research On Online Learning Recommendation Based On User Behavior Sequence

Online learning has become an important part of modern education.With the exponential growth of learning resources,users find it difficult to satisfy their own needs through simple searches.Online learning recommendation systems have emerged as a solution to this problem.Learning recommendation systems can use users’ historical behaviors to mine their learning intentions and personalize course recommendations.Existing online learning recommendation research is mainly divided into traditional methods such as collaborative filtering and deep learning-based methods.They typically extract static learning interests and do not consider the interference of "interest drift" caused by inconsistent interests during the learning process.Additionally,traditional recommendation modeling often discards features with low frequency of occurrence in interaction information,and does not explore co-occurrence characteristics among items at a finer granularity.Moreover,due to high computational costs,most recommendation models based on user behavior sequences ignore the construction of long-term learning preferences.In response to the above-mentioned problems,this thesis conducted research on online learning recommendation and proposed two recommendation methods based on user behavior sequences.Firstly,a course recommendation model,LRIE,is designed based on the evolution of user learning interests,with the aim of mining user interest evolution and item co-occurrence features.The model constructs a gated recurrent unit hidden layer to extract potential learning interest states in the user behavior sequence,combined with the local activation feature of the attention mechanism and the sequence modeling feature of the gated recurrent unit to filter out noise caused by irrelevant information,and simulated the evolution of user learning interests.Additionally,the model utilizes bilinear feature cross to calculate the correlation of course feature information and extract the co-occurrence relationship between items.Secondly,based on LRIE,a learning recommendation model LST-LR was designed by supplementing long-term user preference information.The model used the Simhash algorithm to model the user’s long-term behavior sequence,encoding the vector of the user behavior sequence and candidate courses using a randomly sampled matrix composed of multiple rounds of hash functions,and forming a small batch digital signature bit by bit.Based on the principle of locality-sensitive hashing,similar projects were quickly aggregated to obtain a representation of the user’s long-term learning preference.On the other hand,LRIE was used to model the user’s short-term behavior sequence,and the user’s long and short-term sequences were processed in parallel.LST-LR dynamically mined user short-term interest evolution and project co-occurrence features,fully integrated long-term learning preferences to represent the user’s learning needs,and effectively reduced the computational complexity in long sequence modeling.The study compared LRIE and LST-LR with related methods on the Canvas Network and MOOCCube datasets,and found that they show superior recommendation performance in modeling short-term and long-term sequences,respectively.Additionally,the study verified that LST-LR can effectively reduce the computational complexity of modeling longterm behavior sequences based on inference time.This thesis focuses on the research of recommendation models based on behavior sequences and provides new ideas for future optimization of online learning recommendations.