Research On Recommendation Algorithm Based On Feature Perception And Memory Perception

With the rise of web2.0,the online behavior patterns of users have changed significantly.Users are no longer merely consumers of information,but also the main producers of information.Their behavioral information like clicks and purchases etc.will be recorded on the Internet.As a result,date of user interactions explodes at an exponential rate which led to the problem of information overload.Recommender systems are an effective way to solve this problem.How to extract valuable information from these rich and diverse data to accurately predict user's preferences will affect platform revenue and user satisfaction directly.Current explorations of most researches focus on feature learning,that is,mining the internal connections between features by learning their commonalities from existing data.However,traditional recommendation algorithms had some limitations and failed to capture the in-depth information between users and items,which makes recommendations with unsatisfactory quality.To address the above problems,this thesis conducts an in-depth research from the perspective of feature engineering,and make innovations based on traditional recommendation algorithms.Through mining the complex relationships from real data at both feature level and field level where features belong to,a better personalized recommendation model is established to improve user satisfaction.The main contributions of this thesis are as follows:1.Traditional item-based collaborative filtering algorithms have the limitations that could only obtain linear relationships between users and items.This thesis proposes an improved collaborative filtering recommendation model which considers different contributions of user historical interactions to the prediction of the targets.Standard attention mechanism cannot work efficiently owing to different lengths of users' historical data,and therefore,this thesis introduced the smoothing parameters to improve this problem.In addition,the residual network is introduced simultaneously to learn nonlinear features to overcome the problem of network degradation in the deep network model and help better the final recommendation results.2.To the problem that the factorization machines failed to generalize unknown features,and different feature interactions are of equal importance to the prediction results,this thesis propose a Feature Perception Factorization Model(FPFM).An attention mechanism is introduced to distinguish the importance of different feature interactions for different prediction targets.FPFM solved the issue of traditional prediction algorithms that would assign a same weight to all feature interactions.It does not require any domain knowledge or artificial feature engineering,and is capable of learning different feature interaction weights from data.By adding a fully connected network,complex patterns are mined from real data by deep learning and high-order features are captured nonlinearly,so that the model has a better expression and generalization ability for complex structural data.3.To solve the issue that the factorization machines only considered the impact of feature-level interactions on predictions,this thesis adopts a new improved idea to propose the Feature-Memory Perceptual Factorization Model(FMPFM).On the basis of FPFM Model,this thesis explicitly considers that the field where features belong to in different instances will affect the feature representations.Hence,the memory perception network is designed to generate a perception factor,and the feature interactions are associated with the fields where they belong to.This further enhances the feature representations of the factorization machines.The fusion of feature-level interactions with field-level interactions as well as the integration of feature engineering,further improve the accuracy of recommendations.In summary,this thesis studies intensively the feature interaction issues in the recommendation process,and makes innovations based on traditional collaborative filtering algorithms by proposing an improved item collaborative filtering model,a feature perception factorization model as well as a feature-memory perceptual factorization model.Various experiments and evaluations are conducted on real public datasets,and comparisons of performance are made with some related and state-of-the-art models.The experiments verify the rationality and effectiveness of the model designed in this thesis,and provide support for further researches on personalized recommendation algorithms in the field of feature interactions.