Database Reference
In-Depth Information
Table 5.1 Simulation results comparing prediction rates
of manual recommendations with those of recommendations
generated by the P-Version
Manual
P-Version
Clicks
4.15
8.63
Baskets
3.57
8.24
Orders
3.71
8.70
Revenue
4.30
8.37
Table 5.2 Prediction rates against the number of transaction steps
Steps
PVs manual
Revenue manual
PVs P-Version
Revenue P-Version
1,000
4.49
3.26
5.28
3.64
2,000
4.37
5.26
5.76
4.90
3,000
4.45
4.93
6.19
5.49
4,000
4.40
4.44
6.54
6.10
5,000
4.36
4.25
6.76
6.56
6,000
4.34
4.27
7.08
6.88
7,000
4.36
4.28
7.23
7.23
8,000
4.32
4.30
7.38
7.38
9,000
4.32
4.40
7.43
7.39
10,000
4.31
4.36
7.45
7.31
11,000
4.26
4.36
7.42
7.44
12,000
4.19
4.37
7.46
7.28
Here, on one hand, the back-then recommendations manually devised by the
shop operators have been issued. As these are static, we obtain the actual prediction
rate of the manual recommendations. On the other hand, we study recommenda-
tions of the P-Version. To render the task more challenging, the algorithm literally
starts from scratch, i.e., the learning starts with the simulation data.
The data are courtesy of a major mail-order company. They have been purified
by removing invalid products as well as multiple calls of product views, which all in
all results in a higher prediction rate. Approximately 600,000 transactions with
4,500 products are left in the purified set. The corresponding prediction rates
(in terms of percentage) are displayed in Table
5.1
.
The P-Version turns out to achieve about twice the rate of the manual recom-
mendations, which gives evidence for the quality and learning performance of the
approach.
As a further refinement, the prediction rates of both types of recommendations
with respect to product views (PVs) and revenue have been simulated for the first
12,000 transaction steps. The results (in terms of percentage) are displayed in
Table
5.2
and Fig.
5.1
.
The P-Version turns out to surpass the manual recommendations in terms of all
measures of prediction quality after approximately 5,000 steps. Moreover, the
learning rate of the former exhibits a logarithmic decay thereafter. After only
