CrossMine

Efficient Classification Across Multiple Database Relations

Authors

• Xiaoxin Yin, Department of Computer Science, UIUC
• Jiawei Han, Department of Computer Science, UIUC
• Jiong Yang, Department of Computer Science, UIUC
• Philip S. Yu, IBM T. J. Watson Research Center

Introduction

CrossMine is a tool for classification in relational databases. The training set is a relational database that contains one target relation, in which every tuple is associated with a class label. The test set is a database of the same schema (and usually all relations remain same except the target relation), but the tuples in the target relation are not labeled. The target relation contains objects to be classified, and the other relations provide useful knowledge for classification. Given a training set, CrossMine builds a rule-based classifier, which can be applied on the test set to predict the class labels of unseen tuples.

Usage Agreement

• Downloading is for internal research usage only.  Redistribution and commercial usage are not permitted. 
• The downloaded software can be only used for performance testing. Please contact the authors for other purposes.

Note

• Feedbacks are always welcome.
• Please contact xyin1@uiuc.edu for bugs and questions about CrossMine.

User Manual

1. Data Format

The training set contains the following files:

• relations_train.des and relaltions_test.des: Description of database schema for the training and testing data (these two files usually only differ in the name of the target relation). Each of them contains the following information. (1) The number and names of relations. (2) All joinable attributes (only keys and foreign-keys need to be listed). (3) The target relation and the target attribute (the class label). Only two values 1 and 2 can be used as class labels. If multi-class classification is to be performed, please build a classifier for each class and select the class with highest score as the predicted class.
• xxx.des and xxx.dat: Description and data of each relation. The description of each relation should contain the names of attributes, their types (nomial or numerical), and their representations (number or string). In the data file, different values are separated by spaces. The first attribute of the target relation must be its IDs, which are distinct for all target tuples. It is suggested that the first attribute of each relation contains the IDs of tuples in that relation.

2. Usage Instructions

1. Run CrossMine.exe.
2. Click "Operation→Read Trainset" and read in "relations_train.des". Now all training data are read in, and you may navigate the data by clicking on relations.
3. Click "Operation→Build Rules" to build rules. Rules are output to "rule.txt".
4. Click "Operation→Read Testset" to read in the testing data.
5. Click "Operation→Classify All Tuples" to perform classification.
6. Click on the ID of any target tuple (either in training or testing set) to see the rule used for classifying this tuple.

Download CrossMine

Downloading is for internal research usage only.  Redistribution and commercial usage are not permitted. Download

References

• Xiaoxin Yin, Jiawei Han, Jiong Yang, and Philip S. Yu, "CrossMine: Efficient Classification across Multiple Database Relations", in Proc.  2004 Int. Conf. on Data Engineering (ICDE'04), Boston, MA, March 2004.
• Xiaoxin Yin, Jiawei Han, Jiong Yang. "Efficient Multi-relational Classification by Tuple ID Propagation" in Workshop on Multi-relational Data Mining in conjunction with KDD 2003, Washington, DC, Aug 2003.