Research On Recommendation Algorithm Based On Implicit Feedback And Multi-Feature Fusion

In recent years,recommendation algorithms based on matrix factorization and those based on Transformer have become hot research directions in the recommendation field.Despite the large number of recommendation algorithms proposed in recent years,these algorithms still have some shortcomings.Firstly,existing matrix factorization-based recommendation algorithms mostly focus on incorporating additional information,neglecting the implicit feedback information that can still be extracted from the user-item cooccurrence matrix.Secondly,existing Transformer-based sequence recommendation algorithms adopt invasive fusion when utilizing heterogeneous auxiliary information,which mixes multiple different vector spaces and leads to information overload and other side effects.In addition,the self-attention layer in Transformers has the flaw of overly focusing on distant items,which limits its ability to capture short-term interest in contextually related items.This thesis addresses the above issues and shortcomings with in-depth research and proposes corresponding recommendation algorithms to solve these problems,which are validated on real-world datasets.The main research contributions of this thesis are as follows:1.To address the shortcomings of existing matrix factorization-based recommendation algorithms,this thesis proposes a Long-term Short-term User Embedding SVD Plus Plus(LSIUE-SVD++)recommendation algorithm that incorporates long-and short-term interest implicit feedback.The algorithm first proposes a method for modeling user longand short-term interest implicit feedback and then incorporates the results of modeling into SVD++.Specifically,the user set composed of implicit feedback on specific item rating behavior is replaced with a user set that has the same long-term and short-term interest preferences.The replacement user set has a more obvious tendency,which is beneficial to simulating the user-item interaction process in matrix factorization algorithms.The experimental results show that LSIUE-SVD++ improves the RMSE by 1.35% and 1.10% on the Movie Lens 100 k and Ciao DVD datasets,respectively,compared with the second-best algorithm,proving the effectiveness of the proposed algorithm.2.To address the shortcomings of existing Transformer-based sequence recommendation algorithms,this thesis proposes a Non-Invasive Long-term Short-term Interest Capturing BERT4Rec(NILSIC-BERT4Rec)algorithm based on capturing user long-term and short-term interest.This algorithm retains the ability of the representation layer and the output of the self-attention layer to be in the same vector space,avoiding issues such as information overload and degradation of the self-attention layer’s expressive power.Additionally,to address the problem of the self-attention layer overly focusing on distant items,this algorithm adds a local encoder to the self-attention layer to strengthen the role of context items in the attention function,enhancing the self-attention layer’s ability to capture local dependencies.Experimental results demonstrate that this algorithm can learn better attention distributions.NILSIC-BERT4 Rec achieves the best experimental results on all datasets,with an average improvement rate of all metrics reaching 7.36%,demonstrating the good performance of NILSIC-BERT4 Rec.