System and method for deep enriched neural networks for time series forecasting
The present teaching relates to method, system, medium, and implementations for machine learning. Upon receiving input data associated with a time series, hidden representations associated with the time series in a feature space are obtained and used to generate a query vector in a query space. Such generated query vector is then used to query relevant historic information related to the time series. The query vector and the relevant historic information are aggregated to generate at least one queried vector, which is aggregated with the hidden representations to generate enriched hidden representations that enhance the expressiveness of the hidden representations.
1. A method implemented on at least one machine including at least one processor, memory, and communication platform capable of connecting to a network for machine learning, the method comprising:
receiving input data associated with a time series;
obtaining hidden representations associated with the time series in a feature space;
generating a query vector based on the hidden representations in a query space, wherein the query vector comprises a forward query vector and a backward query vector, each of which corresponds to a linear transformation of the hidden representations;
querying, based on the query vector and across a plurality of time series, relevant historic information related to the time series, wherein the time series is included in the plurality of time series;
aggregating the relevant historic information with the query vector to generate at least one queried pattern vector; and
enriching the hidden representations by aggregating therewith the at least one queried pattern vector to generate enriched hidden representations, wherein
the enriched hidden representations enhance expressiveness of the hidden representations.
2. The method of claim 1 , wherein the hidden representation corresponds to a set of parameters associated with a model learned for forecasting the time series.
3. The method of claim 1 , further comprising forecasting the time series using the enriched hidden representations based on the input data to generate a prediction.
4. The method of claim 3 , further comprising updating the enriched hidden representations based on an error based loss determined based on a discrepancy between a label of the input data and the prediction.
5. The method of claim 4 , wherein the update to the enriched hidden representations is determined also based on a graph based loss determined based on inter-time-series relationships.
6. The method of claim 1 , wherein
the input data include time series of different types;
the hidden representations are for forecasting the time series of the different types; and
the relevant historic information used to generate the enriched hidden representations encompasses time series of different types.
7. Machine readable and non-transitory medium having information recorded thereon for machine learning, wherein the information, when read by a machine, causes the machine to perform:
receiving input data associated with a time series;
obtaining hidden representations associated with the time series in a feature space;
generating a query vector based on the hidden representations in a query space, wherein the query vector comprises a forward query vector and a backward query vector, each of which corresponds to a linear transformation of the hidden representations;
querying, based on the query vector and across a plurality of time series, relevant historic information related to the time series, wherein the time series is included in the plurality of time series;
aggregating the relevant historic information with the query vector to generate at least one queried pattern vector; and
enriching the hidden representations by aggregating therewith the at least one queried pattern vector to generate enriched hidden representations, wherein
the enriched hidden representations enhance expressiveness of the hidden representations.
8. The medium of claim 7 , wherein the hidden representation corresponds to a set of parameters associated with a model learned for forecasting the time series.
9. The medium of claim 7 , wherein the information, when read by the machine, further causes the machine to perform forecasting the time series using the enriched hidden representations based on the input data to generate a prediction.
10. The medium of claim 9 , wherein the information, when read by the machine, further causes the machine to perform updating the enriched hidden representations based on an error based loss determined based on a discrepancy between a label of the input data and the prediction.
11. The medium of claim 10 , wherein the update to the enriched hidden representations is determined also based on a graph based loss determined based on inter-time-series relationships.
12. The medium of claim 7 , wherein
the input data include time series of different types;
the hidden representations are for forecasting the time series of the different types; and
the relevant historic information used to generate the enriched hidden representations encompasses time series of different types.
13. A system for machine learning, comprising:
a relevant historic information query engine implemented on a processor and configured for
receiving input data associated with a time series,
obtaining hidden representations associated with the time series in a feature space,
generating a query vector based on the hidden representations in a query space, wherein the query vector comprises a forward query vector and a backward query vector, each of which corresponds to a linear transformation of the hidden representations, and
querying, based on the query vector and across a plurality of time series, relevant historic information related to the time series, wherein the time series is included in the plurality of time series;
a graph based historic information aggregator implemented on the processor and configured for aggregating the relevant historic information with the query vector to generate at least one queried pattern vector; and
a feature aggregator implemented on the processor and configured for enriching the hidden representations by aggregating therewith the at least one queried pattern vector to generate enriched hidden representations, wherein
the enriched hidden representations enhance expressiveness of the hidden representations.
14. The system of claim 13 , wherein the hidden representation corresponds to a set of parameters associated with a model learned for forecasting the time series.
15. The system of claim 13 , further comprising a time series forecaster configured for forecasting the time series using the enriched hidden representations based on the input data to generate a prediction.
16. The system of claim 15 , further comprising a global model parameter updater implemented on the processor and configured for updating the enriched hidden representations based on an error based loss determined based on a discrepancy between a label of the input data and the prediction, wherein the update to the enriched hidden representations is determined also based on a graph based loss determined based on inter-time-series relationships.
17. The system of claim 13 , wherein
the input data include time series of different types;
the hidden representations are for forecasting the time series of the different types; and
the relevant historic information used to generate the enriched hidden representations encompasses time series of different types.