Multi-interest Conversational Recommendation Model Based On Implicit Feedbac

In the era of big data,technology has developed at a rapid pace and people have access to a huge amount of information.But when faced with a huge amount of information,it has become a challenge to make the best choice.To alleviate this problem,recommendation systems have emerged.Among recommendation algorithms,the sequential recommendation paradigm plays an important role in recommendation systems.It uses a single vector for the representation of user preferences.However,the user’s session contains multiple interests of the user,and it is difficult to accurately represent the multiple interests of the user with a single vector representation.Therefore,some researchers have devoted their attention to the research of multi-interest modeling,designed a lot of multi-interest models,and achieved good results.Although these multi-interest models have achieved certain results,there are still some problems:(1)Many models let each interest recall a series of candidate items independently,while in real life,the choices made by users at each moment are shared by multiple interests,and users’ multi-interest information is not effectively utilized.(2)Ignoring the problem of interest drift in multi-interest scenarios,most researchers focus on mining users’ multiple interests from their history,trying to better cope with user interest drift by refining users’ different interests.However,when the user’s new interests differ greatly from the historical interests,the correlation between the information in the user’s session and the user’s new interests will be weak.In this case,the information in the user session alone will not be sufficient to support the inference of the new interest,which will lead to a serious degradation of the model performance.(3)When using the display feedback dataset for model training,the model can focus on learning items based on the user’s ratings of the items,thus providing a more accurate representation of the user’s interests.However,in reality,users seldom give their own ratings online,or give random ratings,or even do not give their own ratings.This makes it difficult and costly to train models with implicit feedback datasets.However,little research has been done on this issue.To alleviate the above problems,this paper proposes two models with multiple user interests,and one kind of new loss function.For the first problem,this paper proposes a Multi-I model which mines the multiple interests of the user through a capsule network,and then models the evolutionary path of each interest of the user,giving the representational vector of each interest at the next moment.Finally,through the collaboration of multiple interests,joint decisions are made and recommendations are made for the user.Through extensive experiments it is shown that good results are obtained.For the second problem,the model uses Bert as the underlying architecture,and uses a bidirectional self-attentive mechanism for bidirectional representation of user sessions.This is because user sessions are not strictly ordered in a user interest drift scenario,using a unidirectional representation would damage the performance of the model.We design a relevance information enhancement module that adaptively captures information relevant to users’ new interests from neighbors’ sessions and then adds the relevance information to users’ history sessions,which can effectively enhance the relevance between users’ history sessions and new interests.Meanwhile,we analyze and reveal the impact of the difference in the degree of abstraction between user and neighbor representation vectors during the information fusion of users and neighbors.Further,this paper designs a new attention unit that explicitly models the differential information between users and neighbors,effectively complementing the missing relevance information in user sessions.Through extensive experiments,it is shown that our proposed method obtains a good enhancement effect.For the third problem,this paper firstly summarizes the variation law of the loss of different rating samples in the explicit feedback dataset with model training through experiments,and finds that the model training is highly sensitive to the ratings.Secondly,an adaptive hierarchical training function with dynamic threshold is designed,which can effectively distinguish different rating samples in the dataset and can optimize the implicit feedback dataset into an explicit feedback dataset to a certain extent.Finally,an adaptive hierarchical training strategy is proposed to obtain better training results in the implicit feedback dataset in order to better learn the samples with different ratings.Through extensive experiments,it is shown that the method achieves excellent performance and greatly improves the performance of the model.