Deep Learning Based Approach for Personalized Recommendations

Abstract

Knowledge-graphs are the most effective type of things used by the google to improve and provide the best search result to the user. Knowledge-graphs contain the well-structured information about the users based on entity to user relation. Most of the researcher used the knowledge-graphs to cope with the cold-start and the sparsity-based problems due to its effectiveness. However the techniques which are already proposed mainly rely on manual feature engineering and did not allow the end-to-end training. Similarly, most of the techniques based on homogeneous based KG’s and very few are based on heterogeneous based KG’s. So, there is a need of such technique which not only solve these problems but also help system and model to improve its performance which we will discuss later in depth. Here we propose the deep learning based approach for personalized recommendation with label propagation algorithm which computes the user-item embedding for the particular user which is based on Graph SAGE and trained in the way of GNN like images.Moreover for the experiments In order to know the desired probability we first we construct the knowledgegraph after taking the text file from Microsoft satori and coupled it with the data which is being used to construct the more generic. Knowledge-graph for taking the specific users preferences we first take the user-item embedding for the specific user after applying the scoring function after that we used the labeling algorithm to provide the better labels for the data which is initially unlabeled to provide batter labels so it can be batter represent the neighborhood labels as compared to the baselines. We introduced the label propagation algorithm a semi-supervised learning algorithm which is used for the efficient labelling of the unlabeled data points. By the efficient labelling after we got entity features and results from the GNN model we used it for the efficient labelling to make perfect assumptions. We prove and show our results on the two publicly available data sets namely LAST.FM and movie-lens data sets along with the results we showed our model effectiveness and proved the performance of our method is best as compared to baselines.