High dimensional time series forecasting
Large scale time series forecasting models are described that leverage deep learning. This can include the utilization of temporal convolution networks and long short-term memory (LTSM) units of recurrent neural networks (RNNS). The model architectures can handle very large data sets with a large number of time series. Diverse scaling is provided through use of a scale-free leveling network architecture, and sparse time-series data is managed using a gating approach. A deep temporally regularized matrix factorization approach to time-series forecasting is utilized that can leverage correlations between the time series during both training and prediction.
1. A computer-implemented method, comprising:
obtaining a set of training data including historical data for a set of queries;
training a temporal model using the training data, the training including:
processing the training data using a leveling network to provide normalization of the training data;
processing at least one sparse time series of the set using a gating network; and
analyzing correlations between two or more time series of the set using deep Temporally Regularized Matrix Factorization (TRMF) to determine common patterns in the training data;
receiving a query prefix;
processing, using the trained temporal model, the query prefix to determine a set of related queries and forecast demand values for an upcoming period of time; and
providing, corresponding to the query prefix, a subset of the related queries having highest forecast demand values for suggestions to complete a query starting with the query prefix.
2. The computer-implemented method of claim 1 , wherein the trained temporal model is a temporal convolutional network capable of modeling large scale time-series.
3. The computer-implemented method of claim 1 , wherein the leveling network utilizes a scale-free leveling architecture providing diverse scaling of the training data for different time series.
4. The computer-implemented method of claim 1 , further comprising:
training the temporal model using a stochastic gradient descent (SGD) approach using mini-batches of the training data.
5. The computer-implemented method of claim 1 , wherein the temporal model is a high-capacity long short-term memory (LSTM)-based model.
6. A computer-implemented method, comprising:
obtaining a set of training data including historical demand data for a set of data series;
training a temporal model using the training data, the training including:
processing the training data using a leveling network to provide normalization of the training data; and
processing at least one sparse time series of the set using a gating network; and
determining, using the trained temporal model, a forecast demand for an upcoming period of time, related queries to be suggested based in part on the forecast demand to complete a query starting with a query prefix comprising part of the related queries.
7. The computer-implemented method of claim 6 , further comprising:
analyzing correlations between two or more of the time series of the set using deep Temporally Regularized Matrix Factorization (TRMF) to determine common patterns in the training data for use in the training of the temporal model.
8. The computer-implemented method of claim 6 , further comprising:
training the temporal model using a stochastic gradient descent (SGD) approach using mini-batches of the training data.
9. The computer-implemented method of claim 6 , wherein the temporal model is a temporal convolutional network capable of modeling large scale time-series.
10. The computer-implemented method of claim 6 , wherein the leveling network utilizes a scale-free leveling network architecture providing diverse scaling of the training data for different time series.
11. The computer-implemented method of claim 6 , wherein the historical demand data relates to demand for at least one of an item, a type of content, a type of hardware, a query, a keyword, a set of transactions, or logistical data.
12. The computer-implemented method of claim 6 , wherein the temporal model is a high-capacity long short-term memory (LSTM)-based model.
13. The computer-implemented method of claim 6 , wherein the temporal model is trained to account for non-linear temporal dependencies in the training data.
14. The computer-implemented method of claim 6 , wherein the temporal model utilizes a causal convolutional architecture with dilation.
15. The computer-implemented method of claim 6 , wherein the leveling network determines mean values and deviations for individual time series of the training data.
16. A system, comprising:
at least one processor; and
memory storing instructions that, when executed by the at least one processor, cause the system to:
obtain a set of training data including historical demand data for a set of data series;
train a temporal model using the training data, including:
processing the training data using a leveling network to provide normalization of the training data; and
processing at least one sparse time series of the set using a gating network; and
determine, using the trained temporal model, a forecast demand for an upcoming period of time, related queries to be suggested based in part on the forecast demand to complete a query starting with a query prefix comprising part of the related queries.
17. The system of claim 16 , wherein the instructions when executed further cause the system to:
analyze correlations between two or more of the time series of the set using deep Temporally Regularized Matrix Factorization (TRMF) to determine common patterns in the training data for use in the training of the convolutional model.
18. The system of claim 16 , wherein the instructions when executed further cause the system to:
train the temporal model using a stochastic gradient descent (SGD) approach using mini-batches of the training data.
19. The system of claim 16 , wherein the temporal model is a high-capacity long short-term memory (LSTM)-based model or a temporal convolutional network capable of modeling large scale time-series.
20. The system of claim 16 , wherein the leveling network utilizes a scale-free leveling network architecture providing diverse scaling of the training data for different time series.