Data Mining Techniques

Chapter 9:
Market Basket Analysis and Association Rules
Market basket analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
Market basket data I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Market basket data II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
Association rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
How good is an association rule? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
Building association rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Pizza example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
Calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
Renements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1
Market basket analysis
Undirected data mining technique (no target or response variable).
Kind of problem: what merchandise are customers buying and when?
Which products tend to be purchased together and which are amenable to promotion?
suggest new store layouts;
determine which products to put on special;
indicate when to issue coupons;
tie data to individual customers through a loyalty card or website registration.
Other potential applicationssee p. 288.
Pitfalls:
rules can end up merely describing previous marketing promotions;
product dierentiation may not be apparent until long history.
c Iain Pardoe, 2006 2 / 11
Market basket data I
Point-of-sale transaction data.
Three levels of market basket data:
customers;
orders (purchases, baskets, item sets);
items.
Track customers over time:
average # orders per customer;
average # (unique) items per order;
proportion of customers purchasing a particular product;
average # orders per customer including a particular product;
average quantity ordered per customer including a particular product.
c Iain Pardoe, 2006 3 / 11
2
Market basket data II
Order characteristics.
Item popularity:
in a one-item order;
in a multi-item order;
amongst customers who are repeat purchasers;
over time;
over regions.
Tracking marketing interventions.
Clustering products by usagewhich product(s) in a purchase suggest the purchase of other
particular products at the same time:
association rules (handfuls of items);
cluster analysis (larger sets).
c Iain Pardoe, 2006 4 / 11
Association rules
If a customer buys item A, we expect he/she will also buy item B.
Actionable: useful rules with understandable, high-quality info, e.g., if Barbie then chocolate;
might suggest more prominent product placement (e.g., beer and diaperssee p. 298),
product tie-ins and promotions, or particular ways to advertise products.
Trivial: already known by managers, e.g., if maintenance contract then large appliance;
may reect past marketing or product bundling;
exceptions may signal failures in business operations, data collection, or processing.
Inexplicable: seem to have no explanation and do not suggest a course of action, e.g., if new
hardware store then toilet bowl cleaners;
may reect over-tting.
c Iain Pardoe, 2006 5 / 11
3
How good is an association rule?
If I
1
(condition/antecedent) then I
2
(result/consequent) e.g., if OJ then milk (p. 299, beware
mistakes).
Support = proportion of transactions with I
1
& I
2
, e.g., 1/5 = 20.0%.
Condence = (# transactions with I
1
& I
2
) /
(# transactions with I
1
), e.g., 1/4 = 25.0%.
Lift =condence / proportion of transactions with I
2
, e.g., (1/4) / (1/5) = 1.25:
or, lift = (actual # transactions with I
1
& I
2
) / (expected #trans with I
1
&I
2
if no
relationship) = #(I
1
&I
2
) / (#I
1
#I
2
/ N) = 1 / (4 1 / 5);
excess = # (I
1
& I
2
) (#I
1
#I
2
/ N) =
1 (4 1 / 5) = 0.20.
c Iain Pardoe, 2006 6 / 11
Building association rules
Determine the item set:
select right level of detail (start broad, repeat with ner detail to hone in);
product hierarchies help to generalize items, e.g., frozen food dessert, vegetable, dinner;
hybrid approach depending on price or frequency (analysis easier when roughly same
number of transactions for each item);
virtual items go beyond product hierarchy, e.g., designer labels, low-fat products,
energy-saving options, payment method, day, demographic info, etc.
Calculate counts/probabilities of items and combinations of items.
Analyze support/condence/lift to nd actionable rules.
c Iain Pardoe, 2006 7 / 11
4
Pizza example
c Iain Pardoe, 2006 8 / 11
Calculations
Pr(I
1
&I
2
) Pr(I
1
)
Pr(I1&I2)
Pr(I1)
Pr(I
2
)
Pr(I1&I2)
Pr(I1)Pr(I2)
Rule Support Condence Lift
If M then P 0.25 0.450 0.556 0.425 1.31
If P then M 0.25 0.425 0.588 0.450 1.31
If M then C 0.20 0.450 0.444 0.400 1.11
If C then M 0.20 0.400 0.500 0.450 1.11
If P then C 0.15 0.425 0.353 0.400 0.88
If C then P 0.15 0.400 0.375 0.425 0.88
If (M,P) then C 0.05 0.250 0.200 0.400 0.50
If (M,C) then P 0.05 0.200 0.250 0.425 0.59
If (P,C) then M 0.05 0.150 0.333 0.450 0.74
Best rule: if Pepperoni then Mushroom.
Support 25% and condence 58.8%.
Lift 1.31 means when Pepperoni requested then Mushroom is 31% more likely to be requested
than if Pepperoni & Mushroom were unrelated.
c Iain Pardoe, 2006 9 / 11
5
Renements
Negative rules: when lift < 1 then negating the result produces a better rule, e.g.:
Pr(I
1
&I
2
) Pr(I
1
)
Pr(I1&I2)
Pr(I1)
Pr(I
2
)
Pr(I1&I2)
Pr(I1)Pr(I2)
Rule Support Conf Lift
If (P,C) then not M 0.10 0.150 0.667 0.550 1.21
If (M,P) then not C 0.20 0.250 0.800 0.600 1.33
Overcoming practical limits: nd rules with 2 items, 3 items, etc.
Pruning: reduce # of items and combinations of items considered at each step (e.g., minimum
support threshold).
Problems of large datasets: very computer intensive.
c Iain Pardoe, 2006 10 / 11
Extensions
Using association rules to compare stores, promotions at various times, geographic areas, urban
vs. suburban, seasons, etc. (use virtual items).
Dissociation rules, e.g., if I
1
and not I
2
then I
3
(use sparingly for only the most frequent items).
Sequential analysis using association rules: requires identifying customers and tracking them
over time.
c Iain Pardoe, 2006 11 / 11
6