Introduction next up previous
Next: Example Application: May I Up: Personalizing E-commerce Applications with Previous: Personalizing E-commerce Applications with

1. Introduction

  The advent of e-commerce is forcing radical changes to the landscape of marketing and customer care. Customers are demanding increased flexibility and convenience in accessing information about products, in ordering them, and obtaining service for them. At the same time, businesses are attempting to support (a) ``segment of one" marketing and service to large masses of people [18,20], including intelligent targeted advertising, and intelligent mechanisms to identify and take advantage of profitable and loyal customers; and (b) meaningful dialogues with customers so that quality of service can be improved before customers switch to a competitor. These needs are not restricted to B2C e-commerce; web-sites in B2B e-commerce that are accessed by employees of a business must also provide effective, personalized service. This paper introduces a new approach to personalizing web-based e-commerce sites called DFP (Decision Flow Personalization), that is based on the use of on-line decision support. A central contribution is the use of a novel language for specifying decisions, that supports flowchart constructs and a specialized construct called ``Decision Flow'', that combines rules-based constructs and a variety of specialized constructs to facilitate reasoning based on both heuristics and partial information. The DFP approach is illustrated here by describing the MIHU (May-I-Help-You) prototype system, that proactively offers live assistance to web storefront customers.

A fundamental challenge in supporting personalization through on-line decision support is to create a high-level language for specifying decisions that supports sophisticated reasoning but which at the same time is accessible to business analysts and managers. As a starting point for our work, we interviewed business experts on customer care and personalization to understand the features that they need from a decision support language. The following requirements were determined:

Ability to use both formal (e.g., chaining of rules) and heuristic (e.g., giving scores based on ad hoc combinations of various factors) styles of reasoning;

Ability to use rules where appropriate, and to use flowchart constructs where appropriate;

Ability to work with partial and/or incomplete information;

Possibility for hierarchical, modular structuring;

Ability to bring in outside information (e.g., access to customer profiles, the results of bulk statistical analysis);

Ability to invoke side-effect functions (e.g., database updates, triggering workflows).

A clear and intuitively natural semantics;

A natural correspondence between reports on decisions made and the structure of how the decisions are specified (i.e., primarily the structure of the rule sets); and

The language can be ``owned'' or controlled by business analysts and managers, without relying on programmers that translate the decision specifications into a highly technical format.
Some additional systems requirements were determined:
The on-line decision engine should permit changes to decision specifications with no interruption in service; and

User-friendly authoring of decision policies, including rules.

As detailed in Section 5, the rules-based decision specification languages used in existing approaches for e-commerce personalization (e.g., Manna [15], Blaze [2]) satisfy some but not all of these requirements because of their limited expressive power, and other approaches to decision specification (e.g., expert systems, logic programming) fail to satisfy some of the requirements because they are too rich.

To fill this void we use a new paradigm for specifying decisions, called Vortex. An early version of this paradigm is described in [12], where the focus was on the flexible specification of workflows that incorporated business heuristics. Central to the Vortex paradigm is the notion of ``Decision Flow'', which is a novel combination of rules constructs and workflow-like constructs. As detailed in Section 3, in a Decision Flow the emphasis is on computing attribute values. Some of these are targets of the decision (e.g., should a discount be offered to a customer) and others are intermediate to the decision (e.g., the likelihood that this customer will leave the site before completing the deal). Rules may be used to compute the values of individual attributes, and rules may be used to control what attributes are to be computed. (For example, attributes not relevant to a specific decision can be ignored.) Reports about decisions made can show the values of the target and intermediate attributes, and have structure close to the structure of the Decision Flow.

Decision Flows permit complex reasoning about a broad array of data about web sessions and customers. To take full advantage of this it is important to have access to rich information about the pages a customer is visiting, including the underlying intent of the pages (e.g., is it a catalog page, an instructions page, a shopping cart page) and the content delivered in them (e.g., what is the quality of a search result). Section 4 outlines and compares different approaches for gaining access to that information.

To illustrate the core technology and benefits of DFP this paper introduces the May-I-Help-You (MIHU) prototype system. This system is aimed at reducing the number of abandoned e-commerce transactions. Industry statistics [7] indicate that in the U.S. market, for every online B2C transaction that is completed there are nearly four times as many that are abandoned. Further, 7.8% of the abandoned transactions could be converted into sales by using live Customer Service Representative (CSR) interaction. This translated into $6.1 billion in lost e-commerce sales in 1999, and could lead to a cumulative loss of more than $173 billion in the subsequent 5-year period.

The MIHU system monitors a customer's progress through a web storefront, and uses stored and real-time information to infer values such as the current business value of the session and the frustration level of the customer. The MIHU system can proactively offer the customer discounts or targeted promotions. Further, the MIHU system can offer the customer a ``May I Help You'' window, which invites the customer to interact with a Customer Service Representative (CSR), through text chat, voice chat, and/or escorted web browsing. The decision server accesses both stored information and real-time information, including the current availability of, and load on, the CSRs.

Lucent Technologies is developing a product, called Contact Assist, that will support the functionality of the MIHU prototype system, including both an engine based on Vortex and a flexible mechanism for gathering information from a web server. Contact Assist will be available in mid 2001.

The DFP approach can be used in a wide variety of e-commerce applications involving personalization and customization, including the offering of carefully targeted promotions and discounts, helping with navigation through catalogs or self-help material, guiding a customer through an ordering process, and conducting automated dialogues with the customer. It can also be used in non-commercial web-based applications, including context-aware searching tools and automated customization of portals.

Organization. As noted above, the MIHU system will be used to illustrate the main features of our approach. For this reason, we begin in Section 2 by describing the MIHU system at a high level. Section 3 describes the Vortex and Decision Flow paradigms and illustrates their use in connection with the MIHU system. Section 4 describes approaches for incorporating on-line decision servers, such as Vortex, into web-based e-commerce sites. Section 5 considers related work. Section 6 discusses future research directions.

next up previous
Next: Example Application: May I Up: Personalizing E-commerce Applications with Previous: Personalizing E-commerce Applications with
Rick Hull