Recommendations as a Game: Reinforcement Learning for Recommendation Engines - Realtime Data Mining - page 53

4.4 Verification Methods

Subsequently, we shall address the verification methods for RL procedures that we
shall use in Chap. 5 .
To this end, we employ historical data from real online shops. These consist of
log files that have been generated by the recommendation module of prudsys RDE
(Sect. 12.3.1 ). The files contain basic transactions, as well as the recommendations
issued by the RDE (for the recommendation group). The most relevant columns of
the log files (CSV format) are listed in Table 4.1 (further columns like user ID and
channel ID have been omitted).
The control group serves the purpose of comparing the efficiencies of running
the shop with or without the recommendation engine (Chap. 11 ) .
A row of a log file then looks as follows (for confidentiality reasons, all entries
have been made up):

2009-04-18 11:36:21, 386AC17893,0,0045322,17.48,0,0,7889965
05564556

This row tells us that on April 18, 2009, at 11:36:21 AM, the product 0045322,
the price of which amounts to 17.48 EUR, was viewed within the session
386 AC17893, which belongs to the recommendation group, and the products
7889965 and 05564556 were thereupon recommended.
For the online recommendation algorithms, we make use of the following online
verification method: the historical data are parsed session-wise, and the transactions
are carried out in their original order of sequence. For each product view, the
engine's recommendations are requested. Subsequently, the latter are compared
with the actual views, adds to basket, and purchases and the surveys of correct
predictions as well as forecasted revenue are updated. Table 4.2 illustrates the
procedure by an example of a session.
In step 1, the user views product A and the RE issues the recommendations C and
B. The counter of views is increased by one. In the second step, the user switches to
product D and the RE recommends E and A; the view counter is again increased by

Table 4.1 Description of the columns of the transaction files generated by prudsys RDE
Column label Description

time Date and time of the transaction

transactID Unique ID of the session

group The session's group affiliation (0, recommendation; 1, control group)

itemID Unique ID of the product

price Price of the product

transType Type of transaction (0, click; 1, add to basket; 2, purchase; etc.)

order Number of purchased units (only for transType ¼ 2)

itemsAction Recommended products