Method, system, and computer program product for efficient content- based time series retrieval

A system, comprising:

• at least one processor coupled to a memory and configured to:
• obtain, from at least one database, a plurality of known time series;
• for each known time series of the plurality of known time series:
• compute a pairwise distance matrix between that known time series and each learned template of a plurality of learned templates to generate a plurality of pairwise distance matrices;
• stack the plurality of pairwise distance matrices together to generate a multi-dimensional tensor;
• and
• process, with a residual network, the multi-dimensional tensor by:
• projecting, with a two-dimensional (2D) convolutional layer including an input dimension that includes a plurality of dimensions, the multi-dimensional tensor to a Euclidean space,
• generating, with a rectified linear unit (ReLU) layer, based on the multi-dimensional tensor that is projected to the Euclidean space, an intermediate representation for that known-time series,
• generating, with a plurality of bottleneck building blocks, based on the intermediate representation, a merged representation for that known time-series, wherein each bottleneck building block includes a first convolutional layer that projects the intermediate representation to a bottleneck dimension, a second convolutional layer that further transforms the intermediate representation projected to the bottleneck dimension, and a third convolutional layer that projects the transformed intermediate representation to an output dimension to generate the merged representation for that known time-series,
• generating, with a global pooling layer, based on the merged representation, a multi-dimensional vector for that known time-series, and
• generating, with a linear layer, based on the multi-dimensional vector, a feature vector for that known time series;
• and
• provide the feature vector for each known time series of the plurality of known time series,
• wherein the residual network is trained using a loss function that depends on a batch of training data in which each sample in the batch is a tuple including a query time series, a positive time series, and a negative time series.

A computer program product including a non-transitory computer readable medium including program instructions which, when executed by at least one processor, cause the at least one processor to:

• obtain, from at least one database, a plurality of known time series;
• for each known time series of the plurality of known time series:
• compute a pairwise distance matrix between that known time series and each learned template of a plurality of learned templates to generate a plurality of pairwise distance matrices;
• stack the plurality of pairwise distance matrices together to generate a multi-dimensional tensor;
• and
• process, with a residual network, the multi-dimensional tensor by:
• projecting, with a two-dimensional (2D) convolutional layer including an input dimension that includes a plurality of dimensions, the multi-dimensional tensor to a Euclidean space,
• generating, with a rectified linear unit (ReLU) layer, based on the multi-dimensional tensor that is projected to the Euclidean space, an intermediate representation for that known-time series,
• generating, with a plurality of bottleneck building blocks, based on the intermediate representation, a merged representation for that known time-series, wherein each bottleneck building block includes a first convolutional layer that projects the intermediate representation to a bottleneck dimension, a second convolutional layer that further transforms the intermediate representation projected to the bottleneck dimension, and a third convolutional layer that projects the transformed intermediate representation to an output dimension to generate the merged representation for that known time-series,
• generating, with a global pooling layer, based on the merged representation, a multi-dimensional vector for that known time-series, and generating, with a linear layer, based on the multi-dimensional vector, a feature vector for that known time series;
• and
• provide the feature vector for each known time series of the plurality of known time series,
• wherein the residual network is trained using a loss function that depends on a batch of training data in which each sample in the batch is a tuple including a query time series, a positive time series, and a negative time series.