Embedding calendar knowledge in event-driven inventory forecasting
Systems and methods for automatically forecasting the future availability of one or more resources, such as Internet advertising opportunities, are described herein. In accordance with various embodiments, a forecasting model that accounts for event-driven resource availability is trained based both on historical supply data and calendar information specifying events and event duration. The trained forecasting model is then used to forecast the availability of resources at one or more specified future time periods. In accordance with certain embodiments, the forecasting model comprises a Gaussian process model that has an event-driven kernel as a covariance function.
1. A method for forecasting an availability of resources, comprising:
training, by a processor, a non-parametric forecasting model based both on historical supply data and calendar information,
wherein the historical supply data specifies one or more past time periods in a series of time periods and a number of resources that were available during each of the one or more past time periods,
wherein the calendar information identifies one or more events and, for each event, one or more corresponding time periods in the series of time periods during which the event occurs, thereby producing a trained forecasting model that accounts for event-driven resource availability,
wherein the training of the non-parametric forecasting model includes training of a Gaussian process model,
wherein the training of the Gaussian process model includes training a Gaussian model with an event-driven kernel as a covariance, and
wherein the event-driven kernel includes a distance function that defines a de-correlation level between a calendar feature vector associated with the historical supply data and a calendar feature vector associated with an event type.
receiving, at the processor, a specification of one or more future time periods in the series of time periods; and
generating, by the processor, a forecasted number of resources available for each of the one or more future time periods based on the trained non-parametric forecasting model.
2. The method of claim 1 , wherein the resources comprise online display advertising opportunities.
3. The method of claim 2 , wherein receiving the specification of the one or more future time periods in the series of time periods comprises receiving the specification of the one or more future time periods in the series of time periods from an advertisement booking system, the method further comprising: providing the forecasted number of resources available for each of the one or more future time periods to the advertisement booking system.
4. The method of claim 1 , wherein the event-driven kernel is defined as:
k
(
x
i
,
x
j
)
=
σ
f
2
exp
(
-
1
2
l
2
x
i
-
x
j
2
-
1
2
h
2
α
e
*
(
d
(
x
i
,
e
*
)
-
d
(
x
j
,
e
*
)
)
2
)
+
σ
s
2
δ
x
i
x
j
where
e
*
=
arg
min
e
∈
E
[
d
(
x
i
,
e
)
+
d
(
x
j
,
e
)
]
,
wherein x i and x j comprise calendar feature vectors associated with time periods within the series or time periods, σ f 2 comprises a signal variance, 1 comprises a length scaling parameter, h comprises an event scaling parameter, σ s 2 comprises a noise variance, δ xixj , comprises a Kronecker delta function, E comprises a list of events, d(.) is a distance function that defines a de-correlation level between a calendar feature vector and an event in the list of events, and α e ′>0 comprises a weight associated with event e* in the list of events.
5. The method of claim 1 , further comprising:
generating calendar features for each of the one or more past time periods specified by the historical supply data and for at least one of the one or more time periods identified for each event; and
training the forecasting model based also on the generated calendar features.
6. The method of claim 2 , wherein receiving the specification of the one or more future time periods in the series of time periods comprises receiving the specification of the one or more future time periods in the series of time periods from a system configured to monetize forecasted online advertising opportunities, the method further comprising: providing the forecasted number of resources available for each of the one or more future time periods to the system configured to monetize forecasted online advertising opportunities.
7. The method of claim 2 , further comprising using, by the processor, the forecasted number to support a display advertisement booking system.
8. The method of claim 2 , further comprising using, by the processor, the forecasted number to support a system configured to monetize forecasted online advertising opportunities.
9. An inventory forecasting system, comprising:
a software-implemented training module executing on one or more processor-based computer systems that trains a non-parametric forecasting model based both on historical supply data and calendar information, the historical supply data specifying one or more past time periods in a series of time periods and a number of resources that were available during each of the one or more past time periods, the calendar information identifying one or more events and, for each event, one or more corresponding time periods in the series of time periods during which the event occurs, thereby producing a trained non-parametric forecasting model that accounts for event-driven resource availability, wherein the software implement training module trains non-parametric forecasting model by training a Gaussian process model with an event-driven kernel as a covariance, and wherein a definition of the event driven kernel includes a distance function that defines a de-correlation level between a calendar feature vector associated with the historical supply data and a calendar feature vector associated with an event type; and
a software-implemented forecasting module executing on the one or more processor-based computer systems that receives a specification of one or more future time periods in the series of time periods and generates a forecasted number of resources available for each of the one or more future time periods based on the trained non-parametric forecasting model.
10. The system of claim 9 , wherein the resources comprise online display advertising opportunities.
11. The system of claim 10 , further comprising a software-implemented advertisement booking system executing on the one or more processor-based computer systems that provides the specification of the one or more future time periods in the series of time periods to the forecasting module and that receives the forecasted number of resources available for each of the one or more future time periods from the forecasting module.
12. The system of claim 9 , wherein the event-driven kernel is defined in accordance with:
k
(
x
i
,
x
j
)
=
σ
f
2
exp
(
-
1
2
l
2
x
i
-
x
j
2
-
1
2
h
2
α
e
*
(
d
(
x
i
,
e
*
)
-
d
(
x
j
,
e
*
)
)
2
)
+
σ
s
2
δ
x
i
x
j
where
e
*
=
arg
min
e
∈
E
[
d
(
x
i
,
e
)
+
d
(
x
j
,
e
)
]
,
wherein x i and x j comprise calendar feature vectors associated with time periods within the series or time periods, σ f 2 comprises a signal variance, 1 comprises a length scaling parameter, h comprises an event scaling parameter, σ s 2 comprises a noise variance, δ xixj , comprises a Kronecker delta function, E comprises a list of events, d(.) is a distance function that defines a de-correlation level between a calendar feature vector and an event in the list of events, and α e ′>0 comprises a weight associated with event e* in the list of events.
13. The system of claim 9 , further comprising: a software-implemented calendar feature projection module executing on the one or more processor-based computer systems that generates calendar features for each of the one or more past time periods specified by the historical supply data and for at least one of the one or more time periods identified for each event, wherein the model training module trains the forecasting model based also on the generated calendar features.
14. The system of claim 10 , further comprising a software-implemented system configured to monetize forecasted online advertising opportunities executing on the one or more processor-based computer systems that provides the specification of the one or more future time periods in the series of time periods to the forecasting module and that receives the forecasted number of resources available for each of the one or more future time periods from the forecasting module.
15. The system of claim 10 , further comprising a software-implemented module executing on one or more processor-based computer systems that use the forecasted number to support a display advertisement booking system.
16. The system of claim 10 , further comprising a software-implemented module executing on one or more processor-based computer systems that the forecasted number to support a system configured to monetize forecasted online advertising opportunities.
17. A computer program product comprising a non-transitory computer-readable storage medium having computer program logic recorded thereon for enabling a processing unit to forecast an availability of resources, the computer program logic comprising:
first computer program logic that is executable by the processing unit to receive a specification of one or more future time periods in a series of time periods; and
second computer program logic that is executable by the processing unit to generate a forecasted number of resources available for each of the one or more future time periods based on a non-parametric forecasting model,
wherein the non-parametric forecasting model is trained based both on historical supply data and calendar information, the historical supply data specifying one or more past time periods in the series of time periods and a number of resources that were available during each of the one or more past time periods and the calendar information identifying one or more events and, for each event, one or more corresponding time periods in the series of time periods during which the event occurs, and
wherein the non-parametric forecasting model is trained using a Gaussian Process Model with an event-driven kernel as a covariance and the event driven kernel includes a distance function that defines a de-correlation level between a calendar feature vector associated with the historical supply data and a calendar feature vector associated with an event type.
18. The computer program product of claim 17 , wherein the event-driven kernel is defined as:
k
(
x
i
,
x
j
)
=
σ
f
2
exp
(
-
1
2
l
2
x
i
-
x
j
2
-
1
2
h
2
α
e
*
(
d
(
x
i
,
e
*
)
-
d
(
x
j
,
e
*
)
)
2
)
+
σ
s
2
δ
x
i
x
j
where
e
*
=
arg
min
e
∈
E
[
d
(
x
i
,
e
)
+
d
(
x
j
,
e
)
]
,
wherein x i and x j comprise calendar feature vectors associated with time periods within the series or time periods, σ f 2 comprises a signal variance, 1 comprises a length scaling parameter, h comprises an event scaling parameter, σ s 2 comprises a noise variance, δ xixj , comprises a Kronecker delta function, E comprises a list of events, d(.) is a distance function that defines a de-correlation level between a calendar feature vector and an event in the list of events, and α e ′>0 comprises a weight associated with event e* in the list of events.
19. The computer program product of claim 17 , wherein the resources comprise online display advertising opportunities.
20. The computer program product of claim 19 , wherein the computer program logic further comprises third computer program logic that is executable by the processing unit to use the forecasted number to support a display advertisement booking system.