Bidirectional Tensor Learning Algorithm

Bidirectional relationship is an important issue in brain-like collaborative learning.A variety of social behaviors based on social interactions and perceptions among human individuals require the support of bidirectional relationships.Multi-faceted communication will generate a large amount of user behavior information.In bidirectional relationships,users' behavior information is particularly important.Bidirectional problems such as recommendation,prediction,and interpolation based on this information have been extensively studied,and these information usually have multiple labels.Therefore,the concept of tensor is introduced,and the bidirectional tensor learning algorithm is proposed.The main work of this article is as follows:(1)A context-aware tensor Tucker decomposition algorithm based on matrix factorization is proposed.The algorithm uses the slicing operation to extract the matrix of invisible positions from the tensor,uses matrix decomposition to estimate and fill the value of the invisible positions,and uses the extracted context information to decompose the third-order tensor of the trajectory data into a prediction.The result Shows that the algorithm works better.(2)A Bayesian tensor decomposition interpolation algorithm based on log-normal distribution is proposed.This algorithm iterates a set of random numbers obeying the log-normal distribution,obtains Gibbs sampling model through Markov chain Monte Carlo algorithm(MCMC),and then trains the model to obtain experimental results.The results show that the model can obtain more accurate data interpolation performance in processing third-order tensor data than other methods.(3)A non-negative tensor decomposition interpolation algorithm based on variational Bayes is proposed.The MCMC algorithm using Gibbs sampling is a random approximation method,while variational Bayes is a deterministic approximation method.This algorithm uses the variational Bayesian theory to approximate the posterior distribution of the variables,and adds non-negative constraints to the decomposition of the tensor.The results show that the performance of the algorithm with non-negative constraints is better.In summary,the tensor model can better store the original structure of the data,solve some existing bidirectional problems such as recommendation,prediction,interpolation,etc.,and can solve the dimensional disaster that traditional matrices may produce.Adding Bayesian inference and non-negative constraints of decomposition can improve the performance of tensor decomposition algorithm.