Fig. 1

a Schematic of WormTensor. The neural activity data matrices measured for M animals are transformed into the distance matrices and the membership matrices (binary matrices). In the cluster-based similarity partitioning algorithm (CSPA), the consensus matrix is averaged over all the membership matrices and used for downstream clustering and visualization (t-distributed stochastic neighbor embedding (t-SNE) and uniform manifold approximation and projection (UMAP)). WormTensor, on the other hand, does not take the average, but regards the multiple membership matrices as a third-order tensor, applies tensor decomposition, and uses the computed K dimensional factor matrix U for downstream analysis. WormTensor also generates weight vector W, which contains the weights for data from each of the animals. b Neuronal activities of neurons with the same phase. In this case, the positive correlation coefficient between AVAR and RIMR is maximized when translating AVAR to the right by \(\tau = 35\). c Neuronal activities of neurons with reverse phases. In this case, the negative correlation coefficient between ASEL and ASER is minimized when translating ASEL to the right by \(\tau = 14\). mSBD handles both (b) and (c) cases in a unified manner by taking the absolute value of the correlation coefficient