Boon Toh Low

Model combination in the multiple-data-batches scenario

The approach of combining models learned from multiple batches of data provide an alternative to the common practice of learning one model from all the available data (i.e., the data combination approach). This paper empirically examines the base-line behaviour of the model combination approach in this multiple-data-batches scenario. We find that model combination can lead to better performance even if the disjoint batches of data are drawn randomly from a larger sample, and relate the relative performance of the two approaches to the learning curve of the classifier used. The practical implication of our results is that one should consider using model combination rather than data combination, especially when multiple batches of data for the same task are readily available. Another interesting result is that we empirically show that the near-asymptotic performance of a single model, in some classification task, can be significantly improved by combining multiple models (derived from the same algorithm) if the constituent models are substantially different and there is some regularity in the models to be exploited by the combination method used. Comparisons with known theoretical results are also provided.

Classifying Unseen Cases with Many Missing Values

Handling missing attribute values is an important issue for classifier learning, since missing attribute values in either training data or test (unseen) data affect the prediction accuracy of learned classifiers. In many real KDD applications, attributes with missing values are very common. This paper studies the robustness of four recently developed committee learning techniques, including Boosting, Bagging, Sasc, and SascMB, relative to C4.5 for tolerating missing values in test data. Boosting is found to have a similar level of robustness to C4.5 for tolerating missing values in test data in terms of average error in a representative collection of natural domains under investigation. Bagging performs slightly better than Boosting, while Sasc and SascMB perform better than them in this regard, with SascMB performing best.

An expert advisory system for the tourism industry

Abstract This paper describes a practical experience gained in the development of ANESTA, an expert system for the tourism industry in Hong Kong. Particular emphasis is placed on the formulation of heuristics in knowledge engineering and the design of a user friendly interface. Problems associated with the implementation experience and future directions are also discussed.

Semantic Expectation-Based Causation Knowledge Extraction: A Study on Hong Kong Stock Movement Analysis

Human beings generally analyze information with some kinds of semantic expectations. This not only speeds up the processing time, it also helps to put the analysis in the correct context and perspective. To capitalize on this type of intelligent human behavior, this paper proposes a semantic expectation-based knowledge extraction methodology (SEKE) for extracting causation relations from text. In particular, we study the application of a causation semantic template on the Hong Kong Stock market movement (Hang Seng Index) with English financial news from Reuters, South China Morning Post and Hong Kong Standard. With one-month data input and over a two-month testing period, the system shows that it can correctly analyzes single reason sentences with about 76% precision and 74% recall rates. If partial reason extraction (two out of one reason) is included and weighted by a factor of 0.5, the performance is improved to about 83% and 81% respectively. As the proposed framework is language independent, we expect cross lingual knowledge extraction can work better with this semantic expectation-based framework.

Extracting Causation Knowledge from Natural Language Texts

SEKE2 is a semantic expectation-based knowledge extraction system for extracting causation relations from natural language texts. It is inspired by capitalizing the human behavior of analyzing information with semantic expectations. The framework of SEKE2 consists of different kinds of generic templates organized in a hierarchical fashion. All kinds of templates are domain independent. They are robust and enable flexible changes for different domains and expected semantics. By associating a causation semantic template with a set of sentence templates, SEKE2 can extract causation knowledge from complex sentences without full-fledged syntactic parsing. To demonstrate the flexibility of SEKE2 for different domains, we study the application of causation semantic templates on two domain areas of news stories, namely, Hong Kong stock market movement and global warming.

A Workflow Model for Chinese Business Processes

This paper proposes a new workflow model (MAWM) targeted at Chinese business processing. The cultural dependent characteristics of Chinese business processes are identified and taken into account in the MAWM model. MAWM adopts a decentralized modeling style based on the conventional business perspective on labor division and cooperation. This idea is reflected in the concept of agent-workflow. Activity-based and communication-based modeling methods are integrated in the process model. They are used to model different types of business activities. In addition, the model is formalized using Process Algebra and a simplified algorithm is proposed for safety verification of workflow processes.

Learning from Batched Data: Model Combination Versus Data Combination

Combining models learned from multiple batches of data provide an alternative to the common practice of learning one model from all the available data (i.e. the data combination approach). This paper empirically examines the base-line behavior of the model combination approach in this multiple-data-batches scenario. We find that model combination can lead to better performance even if the disjoint batches of data are drawn randomly from a larger sample, and relate the relative performance of the two approaches to the learning curve of the classifier used. In the beginning of the curve, model combination has higher bias and variance than data combination and thus a higher error rate. As training data increases, model combination has either a lower error rate than or a comparable performance to data combination because the former achieves larger variance reduction. We also show that this result is not sensitive to the methods of model combination employed. Another interesting result is that we empirically show that the near-asymptotic performance of a single model in some classification tasks can be significantly improved by combining multiple models (derived from the same algorithm) in the multiple-data-batches scenario.

Improving the performance of neural networks in classification using fuzzy linear regression

Abstract In this paper, we apply the fuzzy linear regression (FLR) with fuzzy intervals analysis into a neural network classification model. The FLR works as a data handler and separates the data sample into two groups. By training two independent neural works with these two groups, we can better describe the distribution space of the corresponding data sample with two different functions, rather than using only one function. The experimental result shows that our approach improves the accuracy of classification.

An Integrated Approach for Flexible Workflow Modeling

This paper proposes an integrated workflow model known as the Multi-Agent Workflow Model (MAWM). MAWM was designed to offer high flexibility in business process modeling. The central idea of MAWM is based on our conventional business perspective about labor division and cooperation. This idea is reflected in the concept agent-workflow. MAWM includes three sub-models, namely organization, process and data sub-models. Activity-based and communication-based methods are integrated in the process sub-model. They are used to model different types of business activities.

A Note on Prototypes, Convexity and Fuzzy Sets

The work on prototypes in ontologies pioneered by Rosch [10] and elaborated by Lakoff [8] and Freund [3] is related to vagueness in the sense that the more remote an instance is from a prototype the fewer people agree that it is an example of that prototype. An intuitive example is the prototypical “mother”, and it is observed that more specific instances like ”single mother”, “adoptive mother”, “surrogate mother”, etc., are less and less likely to be classified as “mothers” by experimental subjects. From a different direction Gärdenfors [4] provided a persuasive account of natural predicates to resolve paradoxes of induction like Goodman’s “Grue” predicate [5]. Gärdenfors proposed that “quality dimensions” arising from human cognition and perception impose topologies on concepts such that the ones that appear “natural” to us are convex in these topologies. We show that these two cognitive principles — prototypes and predicate convexity — are equivalent to unimodal (convex) fuzzy characteristic functions for sets. Then we examine the case when the fuzzy set characteristic function is not convex, in particular when it is multi-modal. We argue that this is an indication that the fuzzy concept should really be regarded as a super concept in which the decomposed components are subconcepts in an ontological taxonomy.