Multitask transfer learning for optimization of targeted promotional programs
Multitask learning is applied to predict a customer's propensity to purchase an item within a particular category of items. Then, the network is tuned using transfer learning for a specific promotional campaign. Retail revenue and promotional revenue are jointly optimized, conditioned on customer trust. Accordingly, a particular promotional program may be selected that is specific to the user.
1. A computer-implemented method comprising:
training a base model with existing customer records describing purchases within an item collection having a plurality of categories, the existing customer records being associated with items across the plurality of categories, the base model comprising a neural network, wherein the base model is trained on a dataset according to a loss function:
Loss
(
X
,
Y
)
=
1
M
K
∑
u
=
1
M
∑
i
=
1
K
(
f
(
X
u
i
)
-
Y
u
i
)
2
wherein M is a total number of customers, K is a number of categories being predicted, and f(X ui ) is an output of a neural network f for customer u and category i given an input X ui ;
tuning the base model for at least one specific promotion task via a transfer learning process to generate relevancy predictions for one or more of the plurality of categories, the relevancy predictions being specific to a particular user; and
combining the relevancy predictions and advertising revenue to estimate overall performance of the at least one specific promotion task.
2. The computer-implemented method of claim 1 , wherein the base model has an input feature vector X and an output feature matrix Y, wherein each of a plurality of values in the input feature vector X corresponds to a score for a customer associated with a particular product category, and wherein each of a plurality of values in the output feature matrix Y corresponds to a purchase propensity score.
3. The computer-implemented method of claim 2 , wherein the input feature vector X for a particular customer u is given by:
X
u
=
∑
i
=
1
n
w
t
u
,
i
v
p
u
,
i
C
p
u
,
i
wherein C is an encoded matrix representing a product to category membership, v is a vector of prices, where v i represents a price for a particular product, p u,i represents a user's ith interaction with a product, and t u,i represents an index of an interaction type, w t u,i is a vector of product prices, and v p u,i is a vector of interaction weights.
4. The computer-implemented method of claim 1 , wherein tuning the base model comprises removing a final layer of the base model and retraining a new final layer to predict campaign-specific sales.
5. The computer-implemented method of claim 1 , wherein combining the relevancy predictions and advertising revenue to estimate overall revenue comprises optimizing total revenue by selecting one or more advertising campaigns for each customer to balance total revenue.
6. The computer-implemented method of claim 5 , wherein total revenue comprises a sum of retail revenue from a customer and promotional revenue from a sponsor.
7. The computer-implemented method of claim 5 , wherein total revenue is derived from an overall yield equation:
o
u
,
i
=
p
u
,
i
(
1
τ
2
-
1
)
∑
j
y
^
u
,
i
2
∑
j
p
u
,
j
2
+
y
^
u
,
i
τ
wherein ŷ u,i represents a score generated by a second stage transfer model, p u,i is constant for a user u and promotional campaign i, and τ is a selectable constant.
8. The computer-implemented method of claim 1 , wherein the at least one specific promotion task corresponds to a particular brand or particular category from among the plurality of categories.
9. The computer-implemented method of claim 1 , wherein the existing customer records include records of item views, additions of items to a shopping cart, and category spend for each of a plurality of users.
10. The computer-implemented method of claim 1 , further comprising, based on the estimate of overall performance, implementing the at least one specific promotion task.
11. A promotion generation system comprising:
a computing system including a processor operatively coupled to a memory subsystem, the memory subsystem storing customer records and instructions which, when executed, cause the computing system to:
train a base model with the customer records, the customer records describing purchases within an item collection having a plurality of categories and being associated with items across the plurality of categories, the base model comprising a neural network, wherein the base model is trained on a dataset according to a loss function;
Loss
(
X
,
Y
)
=
1
M
K
∑
u
=
1
M
∑
i
=
1
K
(
f
(
X
u
i
)
-
Y
u
i
)
2
wherein M is a total number of customers, K is a number of categories being predicted, and f(X ui ) is an output of a neural network f for customer u and category i given an input X ui ;
tune the base model for at least one specific promotion task via a transfer learning process to generate relevancy predictions for one or more of the plurality of categories, the relevancy predictions being specific to a particular user; and
combine the relevancy predictions and advertising revenue to estimate overall revenue.
12. The promotion generation system of claim 11 , wherein the computing system is further configured to automatically identify one or more recommended promotions and transmit an identifier of the one or more recommended promotions to a retail website.
13. The promotion generation system of claim 11 , wherein the computing system is communicatively connected to a retailer website, and wherein the instructions cause the computing system to:
identify the at least one specific promotion task from among a plurality of different promotion task candidates as an optimized promotion task to be presented to a particular customer based on the customer records; and
automatically present to the particular customer a promotion in accordance with the at least one specific promotion task via the retailer website.
14. The promotion generation system of claim 13 , wherein each of the plurality of different promotion task candidates is assessed relative to at least the particular customer to identify the at least one specific promotion task.
15. The promotion generation system of claim 11 , wherein the computing system is communicatively connected to a retailer website, and wherein the instructions cause the computing system to:
identify the particular customer to present a promotion corresponding to the at least one specific promotion task as an optimized promotion task, the particular customer being identified from among a plurality of customers; and
automatically present to the particular customer a promotion in accordance with the at least one specific promotion task via the retailer website.
16. The promotion generation system of claim 11 , wherein identifying the particular customer comprises optimizing total revenue by selecting one or more advertising campaigns for each customer to balance total revenue.
17. A promotion generation system comprising:
a computing system including a processor operatively coupled to a memory subsystem, the memory subsystem storing customer records and instructions which, when executed, cause the computing system to:
train a base model with the customer records, the customer records describing purchases within an item collection having a plurality of categories and being associated with items across the plurality of categories, the base model comprising a neural network, wherein the base model is trained on a dataset according to a loss function;
Loss
(
X
,
Y
)
=
1
M
K
∑
u
=
1
M
∑
i
=
1
K
(
f
(
X
u
i
)
-
Y
u
i
)
2
wherein M is a total number of customers, K is a number of categories being predicted, and f(X ui ) is an output of a neural networkffor customer u and category i given an input X ui ;
tune the base model for at least one specific promotion task via a transfer learning process to generate relevancy predictions for one or more of the plurality of categories, the relevancy predictions being specific to at least one particular user;
combine the relevancy predictions and advertising revenue to estimate overall revenue associated with the at least one particular user and the at least one specific promotion task; and
transmit an instruction to an online retail platform to automatically present a promotion associated with the at least one specific promotion task to the at least one particular user.
18. The promotion generation system of claim 17 , further comprising the online retail platform, wherein the online retail platform comprises at least one of a retail website server and a mobile application server.
19. The promotion generation system of claim 17 , wherein the customer records are associated with a plurality of customers including the at least one particular user.
20. A computer-implemented method comprising:
training a base model with existing customer records describing purchases within an item collection having a plurality of categories, the existing customer records being associated with items across the plurality of categories, the base model comprising a neural network, wherein:
the base model has an input feature vector X and an output feature matrix Y, wherein each of a plurality of values in the input feature vector X corresponds to a score for a customer associated with a particular product category, and wherein each of a plurality of values in the output feature matrix Y corresponds to a purchase propensity score; and
the input feature vector X for a particular customer u is given by:
X
u
=
∑
i
=
1
n
w
t
u
,
i
v
p
u
,
i
C
p
u
,
i
wherein C is an encoded matrix representing a product to category membership, v is a vector of prices, where v i represents a price for a particular product, p u,i ; represents a user's ith interaction with a product, and t u,i represents an index of an interaction type, w t u,i is a vector of product prices, and v p u,i is a vector of interaction weights;
tuning the base model for at least one specific promotion task via a transfer learning process to generate relevancy predictions for one or more of the plurality of categories, the relevancy predictions being specific to a particular user; and
combining the relevancy predictions and advertising revenue to estimate overall performance of the at least one specific promotion task.
21. A computer-implemented method comprising:
training a base model with existing customer records describing purchases within an item collection having a plurality of categories, the existing customer records being associated with items across the plurality of categories, the base model comprising a neural network;
tuning the base model for at least one specific promotion task via a transfer learning process to generate relevancy predictions for one or more of the plurality of categories, the relevancy predictions being specific to a particular user; and
combining the relevancy predictions and advertising revenue to estimate overall performance of the at least one specific promotion task, wherein combining the relevancy predictions and advertising revenue to estimate overall revenue comprises optimizing total revenue by selecting one or more advertising campaigns for each customer to balance total revenue wherein total revenue is derived from an overall yield equation:
o
u
,
i
=
p
u
,
i
(
1
τ
2
-
1
)
∑
j
y
^
u
,
i
2
∑
j
p
u
,
j
2
+
y
^
u
,
i
τ
wherein ŷ u,i represents a score generated by a second stage transfer model, p u,i is constant for a user u and promotional campaign i, and τ is a selectable constant.