Will Bridewell

Inductive process modeling

Abstract In this paper, we pose a novel research problem for machine learning that involves constructing a process model from continuous data. We claim that casting learned knowledge in terms of processes with associated equations is desirable for scientific and engineering domains, where such notations are commonly used. We also argue that existing induction methods are not well suited to this task, although some techniques hold partial solutions. In response, we describe an approach to learning process models from time-series data and illustrate its behavior in three domains. In closing, we describe open issues in process model induction and encourage other researchers to tackle this important problem.

An interactive environment for the modeling and discovery of scientific knowledge

Existing tools for scientific modeling offer little support for improving models in response to data, whereas computational methods for scientific knowledge discovery provide few opportunities for user input. In this paper, we present a language for stating process models and background knowledge in terms familiar to scientists, along with an interactive environment for knowledge discovery that lets the user construct, edit, and visualize scientific models, use them to make predictions, and revise them to better fit available data. We report initial studies in three domains that illustrate the operation of this environment and the results of a user study carried out with domain scientists. Finally, we discuss related efforts on model formalisms and revision and suggest priorities for additional research.

Two Kinds of Knowledge in Scientific Discovery

Research on computational models of scientific discovery investigates both the induction of descriptive laws and the construction of explanatory models. Although the work in law discovery centers on knowledge-lean approaches to searching a problem space, research on deeper modeling tasks emphasizes the pivotal role of domain knowledge. As an example, our own research on inductive process modeling uses information about candidate processes to explain why variables change over time. However, our experience with IPM, an artificial intelligence system that implements this approach, suggests that process knowledge is insufficient to avoid consideration of implausible models. To this end, the discovery system needs additional knowledge that constrains the model structures. We report on an extended system, SC-IPM, that uses such information to reduce its search through the space of candidates and to produce models that human scientists find more plausible. We also argue that although people carry out less extensive search than SC-IPM, they rely on the same forms of knowledge--processes and constraints--when constructing explanatory models.

Reducing overfitting in process model induction

In this paper, we review the paradigm of inductive process modeling, which uses background knowledge about possible component processes to construct quantitative models of dynamical systems. We note that previous methods for this task tend to overfit the training data, which suggests ensemble learning as a likely response. However, such techniques combine models in ways that reduce comprehensibility, making their output much less accessible to domain scientists. As an alternative, we introduce a new approach that induces a set of process models from different samples of the training data and uses them to guide a final search through the space of model structures. Experiments with synthetic and natural data suggest this method reduces error and decreases the chance of including unnecessary processes in the model. We conclude by discussing related work and suggesting directions for additional research.

Learning declarative bias

In this paper, we introduce an inductive logic programming approach to learning declarative bias. The target learning task is inductive process modeling, which we briefly review. Next we discuss our approach to bias induction while emphasizing predicates that characterize the knowledge and models associated with the HIPM system. We then evaluate how the learned bias affects the space of model structures that HIPM considers and how well it generalizes to other search problems in the same domain. Results indicate that the bias reduces the size of the search space without removing the most accurate structures. In addition, our approach reconstructs known constraints in population dynamics. We conclude the paper by discussing a generalization of the technique to learning bias for inductive logic programming and by noting directions for future work.

Emerging AI & law approaches to automating analysis and retrieval of electronically stored information in discovery proceedings

This article provides an overview of, and thematic justification for, the special issue of the journal of Artificial Intelligence and Law entitled “E-Discovery”. In attempting to define a characteristic “AI & Law” approach to e-discovery, and since a central theme of AI & Law involves computationally modeling legal knowledge, reasoning and decision making, we focus on the theme of representing and reasoning with litigators’ theories or hypotheses about document relevance through a variety of techniques including machine learning. We also identify two emerging techniques for enabling users’ document queries to better express the theories of relevance and connect them to documents: social network analysis and a hypothesis ontology.

Mindreading deception in dialog

This paper considers the problem of detecting deceptive agents in a conversational context. We argue that distinguishing between types of deception is required to generate successful action. This consideration motivates a novel taxonomy of deceptive and ignorant mental states, emphasizing the importance of an ulterior motive when classifying deceptive agents. After illustrating this taxonomy with a sequence of examples, we introduce a Framework for Identifying Deceptive Entities (FIDE) and demonstrate that FIDE has the representational power to distinguish between the members of our taxonomy. We conclude with some conjectures about how FIDE could be used for inference.

Extracting constraints for process modeling

In this paper, we introduce an approach for extracting constraints on process model construction. We begin by clarifying the type of knowledge produced by our method and how one may apply it. Next, we reviewthe task of inductive process modeling, which provides the required data. We then introduce a logical formalismand a computational method for acquiring scientific knowledge from candidate process models. Results suggestthat the learned constraints make sense ecologically and may provide insight into the nature of the modeled domain. We conclude the paper by discussing related and future work.

Learning process models with missing data

In this paper, we review the task of inductive process modeling, which uses domain knowledge to compose explanatory models of continuous dynamic systems. Next we discuss approaches to learning with missing values in time series, noting that these efforts are typically applied for descriptive modeling tasks that use little background knowledge. We also point out that these methods assume that data are missing at random—a condition that may not hold in scientific domains. Using experiments with synthetic and natural data, we compare an expectation maximization approach with one that simply ignores the missing data. Results indicate that expectation maximization leads to more accurate models in most cases, even though its basic assumptions are unmet. We conclude by discussing the implications of our findings along with directions for future work.

Comparison of Semantic Similarity Measures for Application Specific Ontology Pruning

Comparing the effects of one drug to another drug, based on their similarity, is important in clinical research. Ontology-derived measures of drug-drug similarity may help to automate such analyses on large data sets. However, general drug ontologies can contain hierarchical distinctions that are irrelevant to a particular clinical application and thus may lead to inaccurate semantic similarity measures. We propose that ontology pruning be used to remove unneeded concepts so that the resulting ontology better reflects the semantic distinctions of a particular domain. In this paper, we present a novel pruning strategy for drug ontologies. For three clinical domains, we derive previously developed semantic similarity measures for the automatically pruned ontology and the full drug ontology against those for the expert derived ontology. We show that the values of similarity measures based on our pruned approach are closer to those of the expert derived ontology than to those of the full ontology. Our pruning approach thus provides a standardized domain-specific measure of drug-drug similarity for clinical applications.