Determining feature importance and target population in the context of promotion recommendation

Patent metadata

Field	Value
Publication number	US-10546320-B2
Application number	US-201514826819-A
Country	US
Kind code	B2
Filing date	Aug 14, 2015
Priority date	Aug 14, 2015
Publication date	Jan 28, 2020
Grant date	Jan 28, 2020

Abstract

A system, method and computer program product for determining a target population to be subject to a promotion or offer of goods/services. The target population is determined that strongly prefers a given promotion, while also making sure the target population represents a sizable number of consumers such that profits may be maximized. The system provides an output solution listing available promotion options and one or more corresponding target groups based on solving an optimization problem that incorporates, prior obtained most important customer features for each promotion using historical promotion data and statistics measures. The system may automatically initiate a promotion offering to each of said customers by communicating the promotion to the members of the targeted group of people, wherein a percentage of future transactions to which the promotion is offered is expected to exceed a threshold level.

First claim

What is claimed is:

1. A computer-implemented method for generating targeted promotions comprising:
receiving, at a processing unit, historical promotion data offered to plurality of customers,
receiving, at the processing unit, a set of customer features associated with prior transactions associated with previous promotion offerings and acceptances as recorded in a memory storage device;
generating, at the processing unit, a full predictive model using the set of customer features and the historical promotion data, wherein the full predictive model is executable by the processing unit to predict acceptance of promotions from customers when all customer features of the set of customer features are considered;
determining, at the processing unit, an importance score for each of the customer features, wherein determining the importance score for each customer feature comprises:
generating, at the processing unit, a reduced predictive model for the customer feature using the historical promotion data and a subset of the customer features that excludes the customer feature, wherein the reduced predictive model is executable by the processing unit to predict acceptance of promotions from customers when the customer feature is excluded from the subset of customer features;
computing, for a given customer feature, a probability that at least one customer with the given customer feature will accept a given promotion is maximized under the full predictive model, but is not maximized under the reduced predictive model;
determining, at the processing unit, the importance score for the customer feature based on the probability, the full predictive model and the reduced predictive model, wherein the importance score indicates an effect of the customer feature on the full predictive model;
iteratively performing, by the processing unit, the generating of the reduced predictive model and the determining of importance scores, for each customer feature among the set of customer features until an importance score is determined for all customer features;
selecting, by the processing unit, a number of customer features based on the determined importance scores;
constructing, using transaction data relating to the prior transactions, the set of customer features, and the importance scores, an objective function that maximizes a given threshold;
for the given promotion, determining, at the processing unit, a target group of customers for the given promotion by solving an optimization problem based on the selected customer features, wherein:
the given promotion is among a plurality of promotions;
each promotion among the plurality of promotions corresponds to a respective expected acceptance rate;
the optimization problem is being solved based on the objective function to maximize a difference between the target group's expected acceptance rate for the given promotion and the target group's highest expected acceptance rate for any promotion among the plurality of promotions excluding the given promotion, under a constraint that a probability of a percentage of customers among the target group will receive the given promotion meets the given threshold being maximized by the objective function;
receiving, by the processing unit, a query from a customer among the target group of customers, wherein the query relates to the given promotion; and
sending, by the processing unit automatically, the given promotion to the target group of customers.

2. The method of claim 1 , wherein determining the importance scores comprises:
computing, at the processing unit, an importance score of a given customer feature for the given promotion indicating a probability that a customer with the given customer feature will accept the given promotion is maximized, while recommending a different promotion when the given customer feature is ignored.

3. The method of claim 1 , wherein the importance score of each customer feature n where n∈{1.2 . . . N} for a promotion s*∈S, where S is a set of promotion options, is computed by the processing unit according to:
∫1[ s* =arg max s p ( x,s )]1[ s* ≠arg max s p −n ( x,s )]ƒ( x ) dx

wherein p(x,s) representing a probability that a customer with feature x will accept promotion s in which a full model having all customer features for promotion “s” is implemented and p −n (x,s) is the same probability as p(x,s) estimated without feature n; and
wherein ƒ(x) is function representing an empirical probability distribution of customer features according to:
ƒ(x)=(number of transaction records whose customer feature is x)/T, where T is the number of customer transactions.

4. The method of claim 3 , further comprising:
based on the computed importance scores of customer features, selecting by the processing unit the M most important features for promotion s;
for given feature vector x, constructing an M-dimensional vector of important features as x A , and constructing, by the processing unit, an (N−M) dimensional vector of remaining features as x B ; and
constructing, by the processing unit, a function g(x A ) representing a marginal distribution of the M most important features by integrating over the remaining features x B based on the function ƒ(x).

5. The method of claim 4 , further comprising:
constructing, by the processing unit, the objective function according to:

max

l
⁢

⁢
b

,
ub

⁢

∫
x

⁢

(

p
⁡

(

x
,

s
*

)

-

max

s
≠

s
*

⁢

p
⁡

(

x
,
s

)

)

⁢

1
⁡

[

l
⁢

⁢
b

≤

Assignees

• IBM

Inventors

• Elmachtoub Adam N
• Ettl Markus
• Oh Sechan
• Petrik Marek
• Ravi Rajesh K

Classifications

• G06Q30/0255 Primary

  based on user history · CPC title

• G06Q10/067

  Enterprise or organisation modelling · CPC title