KATHY'S DESIGNING HER NEW restaurant's kitchen and is looking to buy several new high-end stoves, refrigerators and the like. As the salesman prepares to organize the bill of sale, he suggests she may want to consider a warranty on the new equipment. For a mere 7.5% of the purchase price, Kathy can have peace of mind for a full four years. She accepts the offer.
Flash forward: The contractor has finished the installation. Now that the kitchen is complete, Kathy wants to take pictures of the room for posterity, so she buys a new digital camera. As she's paying for it she thinks it's a good idea to add an extra memory card for increased storage. Kathy tosses one into her shopping cart.
Clearly, marketers have to realize that cross selling as many different products as possible in both business-to-business and consumer settings can increase a customer's value to an organization. While there are many ways to approach cross selling, a particularly simple one is to become aware of the products or services customers are inclined to purchase together, or even later as add-ons.
Market basket (or association) analysis is a technique that tackles this data-mining problem. In a nutshell, it means looking at what customers put in their real or virtual shopping cart during a visit to your store.
A primary objective of market-basket analysis is to identify appealing relationships among the purchase behavior characteristics in a customer file. Here's an example.
Frank is the manager of Frank Training, an e-commerce site that sells corporate training DVDs. Each transaction is captured in a database consisting of the DVD titles bought by each customer. Additional items such as workbooks, premiums to inspire employee education like mugs or pins, or training manuals also are available. As a result, each record on Frank's customer file corresponds to one complete customer transaction. While this transaction may consist of only one DVD, some could include several DVDs from a single purchase, as well as other merchandise.
Let's examine the eight customers' transactions shown in the table on page 60. (Actually, Frank has millions of transactions recorded in his database — but we'll choose eight to define the market-basket concept most simply.) Of these eight, no one bought more than five items at any given time. Only one customer had one item in the basket.
The analysis will allow Frank to find associations between the items purchased. What can be derived from the data? First, results of association analysis appear in the following formats:
-
X% of the time, A and B are purchased in the same cart.
-
Since people buy A, the chance of buying B is X%.
By looking at the transactions, we might conclude that:
-
Twenty-five percent of the time, DVD 101 turns up in the same cart as DVD 35. (There are eight total transactions, and customer 391 and 716 chose these two items together.)
-
Since people buy DVD 101, there's a 66% probability they'll also choose DVD 35. (Three customers bought DVD 101 — numbers 391, 716 and 881. But only two, 391 and 716, selected this combination.)
Many marketers may believe that this result could easily be accomplished through some basic cross-tab and spreadsheet. However, think again about the number of possible DVDs Frank's company keeps in stock. Add to this the peripheral items available on the Web site. Very quickly, a multitude of combinations and permutations appear that make market-basket analysis particularly attractive.
Several measures are used to gauge the strength of the resulting relationships. The most basic is referred to as the “lift.”
Let's look at Frank Training again, but this time we'll analyze all the transactions in his database — not just the eight used in our earlier example. Across all transactions, it's found that 8% of customers are buying DVD 6. Upon closer inspection of these transactions, Frank finds that among a sub-segment of customers — those that bought DVD 2-24% of them also picked DVD 6. Thus, Frank has two “populations” that have shown interest in DVD 6. The former group refers to all customers who purchased DVD 6. The latter group points to a sub-segment of DVD 2 customers that seem to have a greater interest than the norm in DVD 6. How much greater? By dividing 24% by 8%, we arrive at a lift of 300%.
Now assume Frank does the same analysis, but rather than focusing on DVD 2 he examines the results produced when co-purchasing DVD 35. He'd see that 8% of his entire customer base bought DVD 6, while 7% of the customers that chose DVD 35 added DVD 6. The lift, thanks to the relationship between the two products, was 88%.
A lift of less than 100%, as Frank sees here, can be understood as a negative one, suggesting that people who buy DVD 35 are less likely to buy DVD 6 than one would expect.
In these two examples, Frank is dealing only with two items at a time, and analyzing the potential outcomes. However, it's possible — and typical — to look at more than two items simultaneously.
The knowledge of which products appear in the same basket has a very obvious and useful application. Frank's e-commerce company can intelligently position items that are sold together to be near each other on a particular Web page. Similarly, a brick-and-mortar retailer can situate such items near each other.
Market-basket analysis also offers significant advantages to database marketers. By definition, database DMers store detailed purchase histories of their customers. By doing an association analysis, a manager can present items that are known to sell well with others. Even new customers can be coaxed into buying additional merchandise by offering them “associated” products based on existing customers' behavior.
An attractive aspect of market-basket analysis is its “run by itself” nature. Often, but not always, all products are considered, and the analysis discerns the most critical combinations based on the magnitude of the lift or some other measure. In most cases little maneuvering is required by the analyst.
Other features include the ability to dig still deeper into the results. Frank may find that the association rules differ by SIC code or area of the country. For example, training DVDs popular with metropolitan publishing houses might not be a hit with Midwestern industrial concerns.
While market-basket analysis provides powerful benefits, there are problems that on the surface seem to be a bit more complex. Let's say that besides offering DVDs, Frank's firm sells cameras and other electronic equipment. Previously, Frank studied and understood the relationships of the various items on his firm's Web site. Recall he concluded that DVD 101 was associated with DVD 35. A fair question to ask is which item entered the market basket first — DVD 101 or DVD 35? The order of the purchases can have an impact on a marketer's strategy. If many or all customers chose DVD101 before they picked DVD 35, that fact may provide the retailer with clues about selling certain follow-up products.
To address this sequence issue, a time factor must be incorporated into the analysis. A sequencing approach is very similar to an association tool, but includes time as a critical dimension.
A resulting sequence rule may read something like this: Customers who buy a DVD recorder are 5.5 times more likely to pick up a digital camcorder three to six months after the initial transaction.
Time-related analyses can include periods such as “at a future date,” “within one calendar year” or “by the next weekend.” A warranty of several years' duration for a new appliance might be an applicable factor in the first month after purchase. However, an extension to that warranty may not be relevant until three of four years out.
It's clear one can arrive at many different relationships as the results of market-basket analysis are reviewed — and that can be a problem. Since DMers can sell hundreds of different products, and add and subtract from inventory daily, these association rules frequently get complicated. That means plenty of work for marketers and data analysts alike.
SAM KOSLOWSKY (sam_koslowsky@harte-hanks.com) is vice president for modeling solutions at Harte-Hanks Inc., New York.
What's in That Basket?
Direct marketers can identify purchasing patterns for cross selling by reviewing the items that customers bought together. Here are some shopping baskets for sample company Frank Training.
| Customer No. | Item 1 | Item 2 | Item 3 | Item 4 | Item 5 |
|---|---|---|---|---|---|
| 391 | DVD 101 | DVD 53 | DVD 35 | DVD 16 | 25 blank DVDs |
| 442 | DVD 12 | DVD 35 | DVD 16 | ||
| 290 | DVD 4 | DVD 6 | DVD 35 | 25 blank DVDs | |
| 506 | DVD 1 | ||||
| 716 | DVD 6 | DVD 35 | DVD 102 | DVD 101 | |
| 881 | DVD 101 | DVD 17 | |||
| 899 | Cleaning cloths | DVD organizer | |||
| 1237 | DVD 2 | DVD 16 |
