| Reciprocal recommendation algorithms can be widely applied in online dating,recruitment,teacher-student selection and other scenarios.Its characteristic is that it requires consideration of the reciprocal preferences between objects and subjects in order to achieve good recommendation results.Most existing reciprocal recommendation algorithms adopt the Top-N recommendation strategy,which generate the list of recommendations based on ranking the correlation between users and items.However,this recommendation list becomes ineffective when the algorithm incorrectly predicts the correlation between users and items or users’ preferences have changed over time.At the same time,highly skewed long-tail item distribution is very common in recommendation systems.Recommendation models trained on dataset with such long-tail item distribution easily recommend popular items to users,causing less popular items to be overlooked and lowering the overall success rate of recommendations.In the context of reciprocal recommendations,this phenomenon is even more severe as it can lead to neglecting the preferences of both parties by recommending too many popular items.To address above problems,this thesis designs and implements a reciprocal recommendation system in the context of job seeking and recruitment.The main work contents are as follows:(1)This thesis proposes an alternative reciprocal recommendation algorithm based on generative adversarial network to address the limitations of the Top-N recommendation strategy and the contradiction between users’ dynamically changing interest preference and fixed feature representation.This method uses two sets of generative adversarial networks to learn the preferences of subjects and objects respectively.In the N-th round of recommendation,the attention mechanism is used to adjust unilateral users’ preferences in real time based on the assumption that users are not satisfied with the previous recommendation.At the same time,the two sets of generative adversarial networks are iteratively cross-trained,transferring and sharing reciprocal information,and adjusting reciprocal preferences using the attention mechanism.Finally,the trained generators are used to make recommendations for the subjects and objects respectively.(2)This thesis proposes a neural collaborative filtering based on transferring knowledge from head items to tail items to address the challenge of the long-tail item distribution.This method firstly conducted classified sampling according to the popularity of reciprocal subjects and objects,and then aggregates them into new reciprocal dataset,which enhances the representation of long-tail items and effectively prevents the long-tail item distribution from becoming more severe after aggregation.This method then constructs a neural collaborative filtering model,which combines the reciprocal nearest neighbor features of objects and subjects in the embedding layer,fully exploiting the local feature information of the reciprocal objects and subjects.Meanwhile,curriculum learning and meta-learning are used to transfer knowledge from head items to tail items to improve the quality of representation for long-tail items,thereby improving the overall success rate of recommendations.(3)Finally,based on the algorithms mentioned in this thesis,a reciprocal recommendation system for job seekers and companies was designed and implemented.This system is divided into the job seeker side and the companie side,with main functions including registration,login,home page recommendation and personal home page.Initially,the system uses the alternative reciprocal recommendation algorithm based on generative adversarial network to promote recommendation diversity.As the number of users increases and the long-tail item distribution becomes more severe,the two algorithms are combined to provide personalized recommendations. |
