Download Estimation on KDD Cup 2003

Janez Brank and Jure Leskovec
Jozef Stefan Institute, Ljubljana, Slovenia

One of the tasks on this year's KDD Cup concerned download estimation.

• The dataset is a collection of papers from the high energy physics -- theory area of arXiv.org.
• For some papers, we are told how many times they have been downloaded from the arXiv.org web servers (actually, for each paper, only downloads in the first two months since it was added to the archive are reported; besides, we also get the exact time of each download).
• The task is to predict, for certain other papers, how many times they have been downloaded in the first two months since being added to the archive.

This page contains some information about the work done on this task by Jure Leskovec and myself. 18 teams submitted their entries for this task, and our predictions turned out to be the most accurate.

• Slides (in PowerPoint format) about our methodology and results: shorter version, longer and newer version (presented as a Solomon Seminar, 28 October 2003).
• Our predictions. This is a tab-separated file, with one (paper-ID, prediction) pair per line.
• A table showing the average prediction error of several models based on different representations of papers.
• A detailed technical report about our work. This includes information about the dataset and about all the features we experimented with (even the unpromising ones). Postscript version, PDF version.
• [New!] A short three-page paper about our work, submitted to SIGKDD Explorations. Postscript version, PDF version.
• [New!] The values of various attributes that we used in our experiments, available for all the training and test papers in an easily parsable plain-text format.

The KDD Cup 2003 organizers have also published the actual download counts of the test papers in question.

Janez Brank To my home page.