On the limited memory BFGS method for large scale optimization
Mathematical Programming: Series A and B
Semantic query optimization in Datalog programs (extended abstract)
PODS '95 Proceedings of the fourteenth ACM SIGACT-SIGMOD-SIGART symposium on Principles of database systems
Agglomerative clustering of a search engine query log
Proceedings of the sixth ACM SIGKDD international conference on Knowledge discovery and data mining
Efficient Data Mining for Path Traversal Patterns
IEEE Transactions on Knowledge and Data Engineering
Conditional Random Fields: Probabilistic Models for Segmenting and Labeling Sequence Data
ICML '01 Proceedings of the Eighteenth International Conference on Machine Learning
Efficiently inducing features of conditional random fields
UAI'03 Proceedings of the Nineteenth conference on Uncertainty in Artificial Intelligence
A machine learning approach to identifying database sessions using unlabeled data
DaWaK'05 Proceedings of the 7th international conference on Data Warehousing and Knowledge Discovery
Hi-index | 0.00 |
When gathering information from multiple independent data sources, users will generally pose a sequence of queries to each source, combine (union) or cross-reference (join) the results in order to obtain the information they need. Furthermore, when gathering information, there is a fair bit of trial and error involved, where queries are recursively refined according to the results of a previous query in the sequence. From the point of view of an outside observer, the aim of such a sequence of queries may not be immediately obvious. We investigate the problem of isolating and characterizing subsequences representing coherent information retrieval goals out of a sequence of queries sent by a user to different data sources over a period of time. The problem has two sub-problems: segmenting the sequence into subsequences, each representing a discrete goal; and labeling each query in these subsequences according to how they contribute to the goal. We propose a method in which a discriminative probabilistic model (a Conditional Random Field) is trained with pre-labeled sequences. We have tested the accuracy with which such a model can infer labels and segmentation on novel sequences. Results show that the approach is very accurate ( 95% accuracy) when there are no spurious queries in the sequence and moderately accurate even in the presence of substantial noise (∼70% accuracy when 15% of queries in the sequence are spurious).