Research On Integrated An Attention Mechanism And Document Context-Aware For Collaborative Filtering Rating Prediction

In the recommender system,we use the user and item information in the data as the input vector of the neural network to make prediction tasks.In order to improve prediction accuracy,the model needs to perform feature extraction efficiently.High dimensional feature vectors are obtained from sparse input data for training.However,it is not advisable to simply increase the efficiency of feature extraction by increasing the number of layers of the neural network.This increases the complexity of the model and may not achieve desired results.In deep learning,the attention mechanism is also a way to simulate the way the human brain operates.It re-assigns the weight of the feature vector to improve the feature extraction efficiency.Therefore,the characteristics of the attention mechanism are used to improve the feature extraction efficiency of collaborative filtering for sparse dataset.The main research contents of this thesis are as follows:1.The influence of attention mechanism on the generalized matrix decomposition prediction model is discussed.The influence of the attention mechanism combining with different parts of the model on the prediction accuracy is proposed,and the general law is obtained based on the experimental results and data sets.A number of comparison experiments were performed on the Movie Lens and Amazon instant Video datasets,and the relationship between the dataset and the attention mechanism was derived based on the experimental results.2.The eigenvectors for different properties are discussed.The concatenation method is better than the point multiplication method in terms of vector combination.Compare the effects of the concatenation method and the point multiplication method on the prediction accuracy of the model on the Movie Lens and Amazon Instant Video datasets while fixing other parameters.The experimental results show that the concatenation method is better than the point multiplication method in improving the prediction accuracy.3.A recommendation method based on collaborative filtering for generalized matrix decomposition model combined with attention mechanism is proposed.The method uses a scoring matrix to obtain the feature input of the model.Meanwhile,the description information of the item is used instead of the sparse one-hot vector as input,and the feature vector is extracted by the convolutional neural network.The description document of the item enriches the characteristics of the input data and facilitates the supervised learning of the prediction model.At the same time,combined with the attention mechanism,the weight of the feature vector is re-assigned to improve the feature extraction efficiency.The optimal attention-collaborative filtering model was compared with other good scoring prediction models on the Movie Lens and Amazon instant Video datasets,and the experimental results were superior to other models.