Building or learning semantic parsers has been an interesting approach for creating natural langu... more Building or learning semantic parsers has been an interesting approach for creating natural language interfaces (NLI’s) for databases. Recently, the problem of imperfect precision in an NLI has been brought up as an NLI that might answer a question incorrectly can render it unstable, if not useless. In this paper, an approach based on ensemble learning is proposed to trivially address the problem of unreliability in an NLI due to imperfect precision in the semantic parser in a way that also allows the recall of the NLI to be improved. Experimental results in two real world domains suggested that such an approach can be promising.
The data collected in the science run of LIGO calls for a thorough analysis of the glitches seen ... more The data collected in the science run of LIGO calls for a thorough analysis of the glitches seen in the gravitational wave channels, as well as in the auxiliary and environmental channels. Rapid growth in size and number of available databases requires fast and accurate data mining algorithms for timely glitch analysis. The study presents a new technique in cluster analysis that we call constrained validation clustering (CV clustering) for mining patterns in gravitational wave burst triggers. The approach avoids using Gaussianity assumptions on data distribution, and was shown to outperform a state of the art in clustering -- G-means -- when K, the number of clusters, is unknown (Tang et. al., 08); experimental results suggested that Guassian mixture assumption can be too strong as a machine learning bias in mining gravitational wave data, evidenced by very severe overfitting of data by G-means. Our current focus is on upgrading CV clustering to utilizing random sampling and stochastic optimization techniques. Preliminary results indicate that such an enhancement can potentially bring about a forty fold increase in computational efficiency while suffering minor degrade in model quality. A current future direction is in further improving quality of models learned by the algorithm for making it an effective approach for real LIGO data analysis.
The fifth science run of LIGO (S5) has recently been concluded. The data collected over 2 years o... more The fifth science run of LIGO (S5) has recently been concluded. The data collected over 2 years of the run call for a thorough analysis of the glitches seen in the gravitational wave channels, as well as in the auxiliary and environmental channels. The study presents two new techniques for cluster analysis of gravitational wave burst triggers. Traditional approaches to clustering treat the problem as an optimization problem in an 'open' search space of clustering models. However, this can lead to problems by producing models that over-fit or under-fit the data as the search is stuck on local minima. The new algorithms tackle local minima by putting constraints in the search process. S-means looks at similarity statistics of burst triggers and builds up clusters that have the advantage of avoiding local minima. Constrained validation clustering tackles the problem by constraining the search in the space of clustering models that are 'non-splittable' models in which the centroids of the left and right child of a cluster (after splitting) are nearest to each other; the region of models that either over-fit or under-fit data (i.e. 'splittable' models) can therefore be effectively avoided when assumptions about data are satisfied. These methods are demonstrated by using simulated data. The results on simulated data are promising and the methods are expected to be useful for LIGO S5 data analysis.
The fifth science run of LIGO (S5) has recently been concluded. The data collected over 2 years o... more The fifth science run of LIGO (S5) has recently been concluded. The data collected over 2 years of the run call for a thorough analysis of the glitches seen in the gravitational wave channels, as well as in the auxiliary and environmental channels. The study presents two new techniques for cluster analysis of gravitational wave burst triggers. Traditional approaches to clustering treat the problem as an optimization problem in an 'open' search space of clustering models. However, this can lead to problems by producing models that over-fit or under-fit the data as the search is stuck on local minima. The new algorithms tackle local minima by putting constraints in the search process. S-MEANS looks at similarity statistics of burst triggers and builds up clusters that have the advantage of avoiding local minima. Constrained validation clustering tackles the problem by constraining the search in the space of clustering models that are 'non-splittable' models in which the centroids of the left and right child of a cluster (after splitting) are nearest to each other; the region of models that either over-fit or under-fit data (i.e. 'splittable' models) can therefore be effectively avoided when assumptions about data are satisfied. These methods are demonstrated by using simulated data. The results on simulated data are promising and the methods are expected to be useful for LIGO S5 data analysis. PACS numbers: 95.75.−z, 95.75.Wx, 95.75.Pq
In this paper, we explored a learning approach which combines different learning methods in induc... more In this paper, we explored a learning approach which combines different learning methods in inductive logic programming (ILP) to allow a learner to produce more expressive hypotheses than that of each individual learner. Such a learning approach may be useful when the performance of the task depends on solving a large amount of classification problems and each has its own characteristics which may or may not fit a particular learning method. The task of semantic parser acquisition in two different domains was attempted and preliminary results demonstrated that such an approach is promising.
Abstract The development of natural language interfaces (NLI's) for database... more Abstract The development of natural language interfaces (NLI's) for databases has been a challenging problem in natural language processing (NLP) since the 1970's. The need for NLI's has become more pronounced due to the widespread access to complex databases ...
Link discovery (LD) is an important task in data mining for counter-terrorism and is the focus of... more Link discovery (LD) is an important task in data mining for counter-terrorism and is the focus of DARPA's Evidence Extraction and Link Discovery (EELD) research program. Link discovery concerns the identification of complex relational patterns that indicate potentially threatening activities in large amounts of relational data. Most data-mining methods assume data is in the form of a feature-vector (a single relational table) and cannot handle multi-relational data. Inductive logic programming is a form of relational data mining that discovers rules in first-order logic from multi-relational data. This paper discusses the application of ILP to learning patterns for link discovery.
Building or learning semantic parsers has been an interesting approach for creating natural langu... more Building or learning semantic parsers has been an interesting approach for creating natural language interfaces (NLI’s) for databases. Recently, the problem of imperfect precision in an NLI has been brought up as an NLI that might answer a question incorrectly can render it unstable, if not useless. In this paper, an approach based on ensemble learning is proposed to trivially address the problem of unreliability in an NLI due to imperfect precision in the semantic parser in a way that also allows the recall of the NLI to be improved. Experimental results in two real world domains suggested that such an approach can be promising.
The data collected in the science run of LIGO calls for a thorough analysis of the glitches seen ... more The data collected in the science run of LIGO calls for a thorough analysis of the glitches seen in the gravitational wave channels, as well as in the auxiliary and environmental channels. Rapid growth in size and number of available databases requires fast and accurate data mining algorithms for timely glitch analysis. The study presents a new technique in cluster analysis that we call constrained validation clustering (CV clustering) for mining patterns in gravitational wave burst triggers. The approach avoids using Gaussianity assumptions on data distribution, and was shown to outperform a state of the art in clustering -- G-means -- when K, the number of clusters, is unknown (Tang et. al., 08); experimental results suggested that Guassian mixture assumption can be too strong as a machine learning bias in mining gravitational wave data, evidenced by very severe overfitting of data by G-means. Our current focus is on upgrading CV clustering to utilizing random sampling and stochastic optimization techniques. Preliminary results indicate that such an enhancement can potentially bring about a forty fold increase in computational efficiency while suffering minor degrade in model quality. A current future direction is in further improving quality of models learned by the algorithm for making it an effective approach for real LIGO data analysis.
The fifth science run of LIGO (S5) has recently been concluded. The data collected over 2 years o... more The fifth science run of LIGO (S5) has recently been concluded. The data collected over 2 years of the run call for a thorough analysis of the glitches seen in the gravitational wave channels, as well as in the auxiliary and environmental channels. The study presents two new techniques for cluster analysis of gravitational wave burst triggers. Traditional approaches to clustering treat the problem as an optimization problem in an 'open' search space of clustering models. However, this can lead to problems by producing models that over-fit or under-fit the data as the search is stuck on local minima. The new algorithms tackle local minima by putting constraints in the search process. S-means looks at similarity statistics of burst triggers and builds up clusters that have the advantage of avoiding local minima. Constrained validation clustering tackles the problem by constraining the search in the space of clustering models that are 'non-splittable' models in which the centroids of the left and right child of a cluster (after splitting) are nearest to each other; the region of models that either over-fit or under-fit data (i.e. 'splittable' models) can therefore be effectively avoided when assumptions about data are satisfied. These methods are demonstrated by using simulated data. The results on simulated data are promising and the methods are expected to be useful for LIGO S5 data analysis.
The fifth science run of LIGO (S5) has recently been concluded. The data collected over 2 years o... more The fifth science run of LIGO (S5) has recently been concluded. The data collected over 2 years of the run call for a thorough analysis of the glitches seen in the gravitational wave channels, as well as in the auxiliary and environmental channels. The study presents two new techniques for cluster analysis of gravitational wave burst triggers. Traditional approaches to clustering treat the problem as an optimization problem in an 'open' search space of clustering models. However, this can lead to problems by producing models that over-fit or under-fit the data as the search is stuck on local minima. The new algorithms tackle local minima by putting constraints in the search process. S-MEANS looks at similarity statistics of burst triggers and builds up clusters that have the advantage of avoiding local minima. Constrained validation clustering tackles the problem by constraining the search in the space of clustering models that are 'non-splittable' models in which the centroids of the left and right child of a cluster (after splitting) are nearest to each other; the region of models that either over-fit or under-fit data (i.e. 'splittable' models) can therefore be effectively avoided when assumptions about data are satisfied. These methods are demonstrated by using simulated data. The results on simulated data are promising and the methods are expected to be useful for LIGO S5 data analysis. PACS numbers: 95.75.−z, 95.75.Wx, 95.75.Pq
In this paper, we explored a learning approach which combines different learning methods in induc... more In this paper, we explored a learning approach which combines different learning methods in inductive logic programming (ILP) to allow a learner to produce more expressive hypotheses than that of each individual learner. Such a learning approach may be useful when the performance of the task depends on solving a large amount of classification problems and each has its own characteristics which may or may not fit a particular learning method. The task of semantic parser acquisition in two different domains was attempted and preliminary results demonstrated that such an approach is promising.
Abstract The development of natural language interfaces (NLI's) for database... more Abstract The development of natural language interfaces (NLI's) for databases has been a challenging problem in natural language processing (NLP) since the 1970's. The need for NLI's has become more pronounced due to the widespread access to complex databases ...
Link discovery (LD) is an important task in data mining for counter-terrorism and is the focus of... more Link discovery (LD) is an important task in data mining for counter-terrorism and is the focus of DARPA's Evidence Extraction and Link Discovery (EELD) research program. Link discovery concerns the identification of complex relational patterns that indicate potentially threatening activities in large amounts of relational data. Most data-mining methods assume data is in the form of a feature-vector (a single relational table) and cannot handle multi-relational data. Inductive logic programming is a form of relational data mining that discovers rules in first-order logic from multi-relational data. This paper discusses the application of ILP to learning patterns for link discovery.
Uploads
Papers by Lappoon Tang