The Research On Dynamic Model Of News Recommendation For One Shot Learning

Personalized recommendations can be pushed to the information we want to know,accelerating people's access to information.Because user interest models in the field of news recommendation tend to fluctuate with changes in behavior information,traditional news recommendation algorithms are not suitable for interest recommendation based on one-shot learning data.Therefore,how to use one-shot learning data to build user interest models and reduce the data complexity in text classification training has become the key to personalized news recommendation.The development of neural network technology provides a new opportunity for the research of news recommendation model oriented to one-shot learnings.The one-shot learning set is characterized by a small sample size and a high feature dimension.In neural network training,if its feature selection is adopted directly,the loss function of the algorithm will tend to show a oscillating downward trend,resulting in overfitting,and the performance of the obtained user interest model is also unstable.The accuracy and recall rate of the main indicators used to evaluate the recommendation results are directly related to the quality of the text classification model,and the strength of the distinction between the feature items between classes and the dependency of the feature items within the class are the keys to define the model.Focusing on the above issues,the thesis mainly did the following work:?1?In order to avoid the over-fitting of features caused by the oscillating decline of the loss function in the user model due to the one-shot learning,a Bi-PSO algorithm based on neural network is proposed.The algorithm uses a forward search method and assigns weights according to the candidate features based on their relationship with the selected features?redundancy or dependence?;using the PSO idea for multi-objective optimization problems,limiting the values of learning factors(8₁ and(8₂.The feature weights and bias values are added to the optimal feature subset as the result of the current training,and used to adjust the balance weight of each candidate feature in order to obtain the final optimized feature set.The simulated experimental data used 5 test data sets of different sizes from the UCI machine learning repository.The experimental results show that the Bi-PSO algorithm can show good classification accuracy in one-shot learning with fewer feature numbers.?2?The weight of a single feature in a text classification model is difficult to characterize its relative importance in different contexts.Therefore,a feature binary weighting method based on feature location and category is proposed.Firstly,the character embedding is obtained by studying the convolution calculation of the convolutional neural network,and the features of different categories and different positions are weighted to construct a feature-weighted TC-Fbfw semantic model.Secondly,in order to improve the training efficiency of word vectors,the TF-IDF term document word embedding is multiplied by the corresponding weights to obtain multiple weighted word embeddings of the sentence,so as to realize the incremental training of the data.Finally,in order to avoid the monotonicity of feature weights,the embedding matrix is used as a channel in the multi-channel CNN model,and the filter is applied to the embedding matrix corresponding to each channel,and then the features of different categories and different positions are subjected to different weighted processing,and input it to CNN for classification.In order to verify the effectiveness of the method,a public Word2Vec5 vector was used as experimental data to conduct a comparative test.The experimental results show that the classification accuracy of the TC-Fbfw semantic model is improved by an average of 1.637%compared with the support vector machine model and decision tree.?3?The design of the collaborative filtering algorithm based on the Bi-PSO algorithm and the TC-Fbfw semantic model is completed,and the timeliness ranking of the news recommendation list is realized based on the context information read by the user.Using the Gensim natural language training tool,the prototype system and related functional modules are designed and implemented.Some user data sets are randomly selected from a well-known domestic news website.Simulation experiments are performed to verify the recommended performance of the recommended model proposed in this thesis.