Database Reference
In-Depth Information
Table 7.4 Results for the ndcHArd data set
Number
of points
Training
correctness (%)
Training
correctness (%)
Total
time [s]
Data matrix
time [s]
Number
of iterations
Level 1
20,000
86.2
84.2
6.3
0.9
45
200,000
85.1
84.8
16.2
8.7
51
2 million
84.9
84.9
114.9
84.9
53
Level 2
20,000
85.1
83.8
134.6
10.3
566
200,000
84.5
84.2
252.3
98.2
625
2 million
84.3
84.2
1,332.2
966.6
668
Table 7.5 Results for the web shop data set with a block size of 10,000 data points
Blocks
Correctness (%)
Total time [s]
Blocks
Correctness (%)
Total time [s]
1
44.79
12.05
10
62.94
96.22
2
53.61
20.52
11
63.32
106.82
3
57.53
29.06
12
63.54
117.51
4
59.29
38.33
13
63.59
128.61
5
60.71
47.50
14
63.88
139.98
6
61.60
56.76
15
64.15
151.32
7
61.97
66.13
16
64.29
163.04
8
62.36
75.83
17
64.51
174.88
9
62.72
85.78
18
64.67
184.72
Example 7.5
Now we will compare the results of the offline Algorithm 7.1 with
that of Algorithm 7.3, its adaptive counterpart. We use a data set suitable for
scoring-based recommendations as they have been described in the introduction,
especially in Example 7.1. The data is from a large web shop and each data
point represents a certain transaction in a web session, such as a click, basket
event, or order. The data is from 1 day. There are 15 attributes from three data
sources:
• Session specific: number of clicks and basket products in current session
• User specific: gender, customer value, etc. (only for recognized users)
• External from host: step in checkout process, availability of products, etc.
As in Example 7.1, the target attribute is 0 if no order was placed within the
session and 1 if something was ordered. The task is to find a good classification
function that predicts at each step of a web session if the user will finally place an
order.
Thus, the data set has 15 dimensions. The number of data points is 177,907. For
simplicity, we use the training set as test set, too. As in the previous example, we
again apply simplicial basis functions which are computationally cheaper the
tensor-product ones. By using the offline Algorithm 7.1, for level 1 we obtain a
training correctness of 67.90 % and the run time is 283 s. Now we apply the
adaptive Algorithm 7.3 to the same problem. For a block size
M ¼
10,000, the
results are shown in Table
7.5
.
