SYSTEM AND METHOD FOR SEQUENTIAL DECISION MAKING FOR CUSTOMER RELATIONSHIP MANAGEMENT

• US 20080249844A1
• Filed: 06/17/2008
• Published: 10/09/2008
• Est. Priority Date: 07/19/2002
• Status: Active Grant

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Abe, Naoki, Pednault, Edwin P.D.
• Granted Patent: US 8,285,581 B2
• Time in Patent Office: Days
• US Class Current: 705/10
• CPC Class Codes: G06Q 10/06375 Prediction of business process outcome or impact based on a proposed change
• CPC Class Codes: G06Q 30/02 Marketing; Price estimation or determination; Fundraising
• CPC Class Codes: G06Q 30/0201 Market modelling; Market analysis; Collecting market data
• CPC Class Codes: G06Q 30/0202 Market predictions or forecasting for commercial activities
• CPC Class Codes: G06Q 30/0204 Market segmentation
• CPC Class Codes: G06Q 30/0226 Incentive systems for frequent usage, e.g. frequent flyer miles programs or point systems
• CPC Class Codes: G06Q 30/0234 Rebates after completed purchase
• CPC Class Codes: G06Q 30/0239 Online discounts or incentives

Abstract

A system and method for sequential decision-making for customer relationship management includes providing customer data including stimulus-response history data, and automatically generating actionable rules based on the customer data. Further, automatically generating actionable rules may include estimating a value function using reinforcement learning.

Claims

1. A method for sequential decision making for customer relationship management, comprising:

providing customer data comprising stimulus-response history data for a population of customers, said stimulus response history data being derived from event data for said customers;

automatically generating actionable rules for optimizing a sequence of decisions over a period of time based on said stimulus-response history data;

estimating a value function using batch reinforcement learning with function approximation, said function approximation representing the value function as a function of state features and actions; and

transforming an output of a value function estimation into said actionable rules.

2. The method according to claim 1, wherein said estimating said value function using batch reinforcement learning with function approximation comprises:

estimating a function approximation of the value function of a Markov Decision Process underlying said stimulus-response history data for said population of customers.

3. The method according to claim 2, wherein said estimating said value function using batch reinforcement learning with function approximation further comprises:

iteratively applying a regression model to training data comprising sequences of states, actions and rewards resulting for said population of customers, and updating in each iteration a target reward value for each state-action pair.

4. The method according to claim 3, wherein said actionable rules specify an action to take given a set of feature values corresponding to a customer, and the action taken corresponding to an action having an approximate maximum value according to said value function for the given set of feature values.

5. The method according to claim 1, wherein said reinforcement learning comprises maximizing a net present value of profits with respect to a customer.

6. The method according to claim 1, wherein said reinforcement learning maximizes a cumulative reward, R, given by

7. The method according to claim 6, wherein γ

comprises a discount factor for calculating a net present value of future rewards based on a given interest rate.

8. The method according to claim 1, wherein said batch reinforcement learning comprises batch reinforcement learning with function approximation based on one of Q-learning and sarsa-learning.

9. The method according to claim 1, wherein said customer data comprises one of consumer data, client data and donor data.

10. The method according to claim 1, wherein said transforming comprises generating training data using a model of said value function, said training data comprising a set of feature values and an action that corresponds to a maximum value for said set of feature values, and outputting a set of rules that are based on said training data, and

wherein said set of rules are generated using one of a classifier and a supervised learning tool.

11. The method according to claim 1, wherein said providing customer data comprises selecting customer data.

12. The method according to claim 11, wherein a value estimation repeatedly calls a data selection module one or more times per each iteration of a value iteration.

13. The method according to claim 12, wherein said value estimation repeatedly calls said data selection module by using a selection criterion to select a restricted subset for use in a value function estimation.

14. The method according to claim 13, wherein said using a selection criterion comprises selecting only those events such that their “

next”

states satisfy a condition that an action taken corresponds to an action with a maximum estimated cumulative value, with respect to an estimated value function from a previous value iteration.

15. The method according to claim 14, wherein said using a selection criteria comprises selecting only those events such that their next n states satisfy a condition that an action taken in each state corresponds to an action with a maximum estimated cumulative value.

16. The method according to claim 3, wherein said stimulus-response history data for each customer comprises a sequence of said customer'"'"'s demographic features and event features, at multiple time points, which collectively reflect the state of said customer at a point in time, an action taken at or around that time by an entity, a response taken by said customer at or around that time, and an amount of profit realized by said entity and associated with said action.

17. The method according to claim 16, wherein said updating said target reward value for each state-action pair uses one of a look-ahead update formula of reinforcement learning, in terms of the reward value corresponding to said state-action pair in the training data, and an estimated value of the value function estimate from a previous iteration, for state-action pairs at a future time following said state-action pair.

18. A system for generating targeted marketing rules for a customer, said system comprising:

a transforming unit for transforming customer transaction data to create derived features;

a data development unit for using said derived features to develop current customer profile data and combined historical customer profile and stimulus response data;

a data mining unit for performing data mining on the combined data to develop a stimulus-response model;

a stimulus optimization unit for performing stimulus optimization using said combined historical customer profile and stimulus-response data and said stimulus-response model with business rules; and

a rule generator for generating customer relationship management (CRM) rules by performing data mining on said combined data and said stimulus optimization.

19. A programmable storage medium tangibly embodying a program of machine-readable instructions executable by a digital processing apparatus to perform a method for sequential decision-making method for customer relationship management, said method comprising:

providing customer data comprising stimulus-response history data for a population of customers, said stimulus response history data being derived from event data for said customers;

automatically generating actionable rules for optimizing a sequence of decisions over a period of time based on said stimulus-response history data;

estimating a value function using batch reinforcement learning with function approximation, said function approximation representing the value function as a function of state features and actions; and

transforming an output of a value function estimation into said actionable rules.