Research On Personalized Recommendation Methods Based On Social Perception

Personalization is becoming a key research issue in the field of recommendation.How to incorporate personal preferences into the recommendation model is a difficult problem to be solved.Previous research often required users to express their preferences for social entities in some way,which was time-consuming and labor intensive.In this paper,inspired by the observations of human cognition and behavior interaction,we suggest modeling the user's preferences from the preference behavior on the social media platform.In this way,we can achieve personalized recommendation without any additional burden on users and it can serve the masses on the social media platform.In order to achieve the goal of personalized aesthetic image recommendation,we collected a series of professional photos and user interaction data.We consider user preferences and general aesthetic standards to deal with the unreliability of user preferences.In addition,we optimize pairwise rankings among entities to alleviate data sparsity,which follows the idea of collaborative filtering.Finally,we develop a novel deep neural network architecture for personalized recommendation modeling.Although user preferences are not clearly known,a lot of experiments have been carried out on two benchmark datasets.The results demonstrate the potential of our personalized recommendation method.Many machine learning applications benefit from multi-task learning.In practical tasks,we usually focus on optimizing a single task goal which usually achieves acceptable performance for this task.But we may overlook some relevant information which may help us to do better on this.We use the multi-task learning approach in personalized social entity recommendations to make the model's generalization performance better.Specifically,we use an MTL CNN structure that implements automatic feature fusion at each layer.The following are the details of feature fusion.Firstly connect the feature graph of different tasks in the feature channel.Then use 1× 1 convolution and batch reduction to combine discriminant features and reduce the dimension.Multitasking learning is the problem of optimizing models for multiple objectives.Our method of combining multi-objective losses is to calculate the linear combination of all task losses.We carefully adjust the weights of different tasks to balance the loss of each task.Different tasks have different learning difficulties.Simple tasks take less time to train than complex tasks.In this paper,we use the early stop mechanism to judge the learning progress of each task.Freeze the learning process for a single task where performance is no longer improved.This can reduce unnecessary calculations to speed up training and reduce overfitting.Our approach benefits from different tasks and achieves superior performance in personalized social entity recommendations compared to learning each task individually.