Recognising innovative companies by using a diversified stacked generalisation method for website classification – the raw results

Introduction

The classification models were trained out by using the Classification and Regression Training package (caret) [1]. The models' parameters were fine-tuned by the 10-fold cross-validation procedure [2].

Cluster parameters

Most computations were carried out on a cluster having the following parameters:

• GPU: NVIDIA Tesla P100;
• CPU: 2.0 GHz Intel® Xeon® Platinum 8167M;
• The number of GPUs: 2;
• The number of CPU cores: 28;
• The number of CPU threads: 56;
• RAM: 192 GB;
• Storage: 3 TB.

Only one model (k-nn) was calculated on a cluster having the following parameters:

• Processor: Intel(R) Core(TM) i7-4770 CPU @ 3.40GHz 3.40 GHz;
• RAM: 16 GB;
• Windows 64 bit.

Performance statistics

All performance statistics are stored in cvs files. Each file corresponds to a particular machine learning method such as a file, "methodName-stat.csv" contains all data regarding a method, "methodName." All files cover the following columns:

• dataSetName – a name of a data set on which evaluation was carried out; there are three possible values: (i) firstPages refers to the first data set (L_D) that contains textual description of a company; (ii)  firstPageLabels refers to the second data set (L_L) that involves link labels that were extracted from an index page; (iii) aggregateDocument refers to the third data set (L_B) that consists of a so-called big document;
• fmeasure - the number of features that were taken into account during  evaluation;
• method  - the name of function in the caret package;
• parameters - the values of parameters received from a tuning phase of a given classification method;
• precision – the value of method’s precision;
• recall – the value of method’s recall;
• fmeasure  - the value of method’s F-measure; 
• error - the value of method’s error;
• acc – the value of method’s.

Time processing statistics

All time processing statistics, like the performance statistics, are stored in cvs files. Each file corresponds to a particular machine learning method such as a file, "methodName-time.csv". All files cover the following columns:

• dataSetName – a name of a data set on which evaluation was carried out; there are three possible values: (i) firstPages refers to the first data set (L_D) that contains textual description of a company; (ii)  firstPageLabels refers to the second data set (L_L) that involves link labels that were extracted from an index page; (iii) aggregateDocument refers to the third data set (L_B) that consists of a so-called big document;
• featureNo - the number of features that were taken into account during  evaluation;
• method  - the name of function in the caret package;
• user - user time elapsed for executing a method as an R process;
• system  - system time elapsed for executing a method as an R process;
• elapsed - total time elapsed for executing a method as an R process.

For more information about user, system and total elapsed time, please see documentation [3].

References

[1] https://cran.r-project.org/web/packages/caret/

[2] https://topepo.github.io/caret/model-training-and-tuning.html

[3] https://stat.ethz.ch/R-manual/R-devel/library/base/html/proc.time.htm