Transcript Propensity to buy

Business System Analysis &
Decision Making
– Data
Mining and Web Mining
Zhangxi Lin
ISQS 5340
Summer II 2006
1
Outline
 Estimating the propensity to buy
 Online recommendation
2
Estimating the Propensity
to Buy
3
Goals for Building Propensity-toBuy Models



Use propensity-to-buy scores for
personalization and to influence dynamic Web
content
Test interventions that are intended to increase
the probability of the user making a purchase
Use propensity-to-buy models to evaluate
banner ads with respect to increasing the
propensity to buy at the target Web site
4
Issues Related to Propensity-toBuy Models



Web log data alone is insufficient to build
these models.
Implementation issues and the use of dynamic
inputs need to be resolved before deciding on
candidate models.
Binomial response modeling is required, but
as an alternative, you can predict the amount
spent in the current session, or you can
employ a two-stage model that predicts both
propensity to buy and amount spent.
5
6
Variable Definitions
7
Variable definitions (Cont’d)
8
SAS Enterprise Miner Model
Dataset
Dataset in
partitioned
Modeling in a
Decision tree
Assess the
outcome
9
Decision Tree
10
Lift Value
11
Captured Response
12
Improving a Model





Add more inputs that may influence the target.
Experiment with transforming inputs.
Combine predictions using an average of two
or more models.
Oversample when the success rate is small,
as it is in the propensity-to-buy example.
Incorporate Bayesian concepts into the
modeling process: prior distributions, profit
and loss matrices (use the target profile
tab/option in Enterprise Miner).
13
Online Recommendation
14
Statement of the Problem




There are M judges and N items.
Each judge j will have assigned a rating R(j,k)
to item k if the judge has evaluated item k,
otherwise the rating will be missing.
For all of the missing ratings, estimate the
rating, or score item k for judge j.
Use the scores to estimate the top S scoring
items that each judge has not rated.
15
Netflix™Movie Rating
16
17
Applications of Recommender
Systems



For customers visiting a retail Web site, use information
from previous purchases to recommend
 Books
 Music CDs
 Movies
An “intelligent” music player: plays music specifically
selected by user, when music has finished and user has
not made a selection in over L seconds, the player
makes a selection for the user based on previous
selections the user has made.
A news service that provides a personalized custom
virtual newspaper to the subscriber based on past news
article preferences. (These are usually content-based
continued...
rather than collaborative.)
18
Applications of Recommender
Systems



Personalize a user’s home page with “interesting”
links, with links based on a recommender system
algorithm that recommends links that should be
interesting to the user.
Send a robot out looking for specific information, score
each Web page using a recommender algorithm, and
then return the K most interesting Web pages sorted
by descending score (search engine applications).
Index a library of information based on recommender
system scores.
19
Issues


Some applications will not have ratings, but
rather 0/1 or No/Yes settings, for example,
Yes, the customer has purchased the title, or
No, the customer has not purchased the title.
Customer preferences may change over time,
or a customer may discover a new artist, so
re-training at regular intervals will be required
for many applications.
20
A Conditional Frequency
Approach






This method may be the current favorite, but it is
generally considered to be inferior to more
sophisticated approaches.
Obtain a subset of your recommender data that only
has customers who purchased the given item, call it
item K. Ratings are 0 (did not buy) or 1 (bought).
Derive the frequency distribution by all other titles.
Sort from highest frequency to lowest.
Recommend a set number of the highest frequency
items. Most Web sites present the top 3 or the top 5
most frequent items.
Example: Amazon.com
21