International Journal on Artificial Intelligence Tools

Choosing the optimal number of clusters is a key issue in cluster analysis. Especially when dealing with more spatial clustering, things tend to be more complicated. Cluster validation helps to determine the appropriate number of clusters present in a dataset. Furthermore, cluster validation evaluates and assesses the results of clustering algorithms. There are numerous methods and techniques for choosing the optimal number of clusters via crisp and fuzzy clustering. In this paper, we introduce a new index for fuzzy clustering to determine the optimal number of clusters. This index is not another metric for calculating compactness or separation among partitions. Instead, the index uses several existing indices to give a degree, or fuzziness, to the optimal number of clusters. In this way, not only do the objects in a fuzzy cluster get a membership value, but the number of clusters to be partitioned is given a value as well. The new index is used in the fuzzy c-means algorithm for the geodemographic segmentation of 285 postal codes.

Dimensionality reduction is a necessary task in data mining when working with high dimensional data. A type of dimensionality reduction is feature selection. Feature selection based on feature ranking has received much attention by researchers. The major reasons are its scalability, ease of use, and fast computation. Feature ranking methods can be divided into different categories and may use different measures for ranking features. Recently, ensemble methods have entered in the field of ranking and achieved more accuracy among others. Accordingly, in this paper a Heterogeneous ensemble based algorithm for feature ranking is proposed. The base ranking methods in this ensemble structure are chosen from different categories like information theoretic, distance based, and statistical methods. The results of the base ranking methods are then fused into a final feature subset by means of genetic algorithm. The diversity of the base methods improves the quality of initial population of the genetic algorithm and thus reducing the convergence time of the genetic algorithm. In most of ranking methods, it's the user's task to determine the threshold for choosing the appropriate subset of features. It is a problem, which may cause the user to try many different values to select a good one. In the proposed algorithm, the difficulty of determining a proper threshold by the user is decreased. The performance of the algorithm is evaluated on four different text datasets and the experimental results show that the proposed method outperforms all other five feature ranking methods used for comparison. One advantage of the proposed method is that it is independent to the classification method used for classification.

Research on networks is increasingly popular in a wide range of machine learning fields, and structural inference of networks is a key problem. Unfortunately, network structural inference is time consuming and there is an increasing need to infer the structure of ever-larger networks. This article presents the Dense Structural Expectation Maximisation (DSEM) algorithm, a novel extension of the well-known SEM algorithm. DSEM increases the efficiency of structural inference by using the time-expensive calculations required in each SEM iteration more efficiently, and can be O(N) times faster than SEM, where N is the size of the network. The article has also combined DSEM with parallelisation and evaluated the impact of these improvements over SEM, individually and combined. The possibility of combining these novel approaches with other research on structural inference is also considered. The contributions also appear to be usable for all kinds of structural inference, and may greatly improve the range, variety and size of problems which can be tractably addressed. Code is freely available online at: http://syntilect.com/cgf/pubs:software.

Learning causal relations from observational data is a fundamental task in knowledge discovery. Recently, an Information Geometric Causal Inference (IGCI) framework is proposed, which represents causal relations by deterministic functions, defines causal mechanisms by a cause-mapping independence postulate and learns causal directions by an information geometry formulation. Observing IGCI's limitation in representing general causal relations and its ambiguity in inferring causal directions, we generalize IGCI's original postulate, and propose a new dependence causal inference (DCI) method where linear correlation and a new cross likelihood (CL) measure are introduced. We prove that CL dependence measure incorporates IGCI as its special case in some sense. Experimental results on synthetic and real-world data verify the effectiveness of our generalization and methods, and show that our method is more robust to disturbances on real data sets.

Many machine learning applications rely on learning distance functions with side information. Most of these distance metric learning approaches learns a Mahalanobis distance. While these approaches may work well when data is in low dimensionality, they become computationally expensive or even infeasible for high dimensional data. In this paper, we propose a novel method of learning nonlinear distance functions with side information while clustering the data. The new semi-supervised clustering approach is called Semi-Supervised Fuzzy clustering with Learnable Cluster dependent Kernels (SS-FLeCK). The proposed algorithm learns the underlying cluster-dependent dissimilarity measure while finding compact clusters in the given data set. The learned dissimilarity is based on a Gaussian kernel function with cluster dependent parameters. This objective function integrates penalty and reward cost functions. These cost functions are weighted by fuzzy membership degrees. Moreover, they use side-information in the form of a small set of constraints on which instances should or should not reside in the same cluster. The proposed algorithm uses only the pairwise relation between the feature vectors. This makes it applicable when similar objects cannot be represented by a single prototype. Using synthetic and real data sets, we show that SS-FLeCK outperforms several other algorithms.

For multiobjective particle swarm optimization (MOPSO), two particles may be incomparable, i. e., not dominated by each other. The personal best and the global best for the particle become less optimal, thus the convergence becomes slow. Even worse, an archive of a limited size can not cover the entire region dominated by the Pareto front, the uncovered region can contain unidentifiable nondominated solutions that are not optimal, and thus the precision the algorithm achieves encounters a plateau. Therefore we propose dimensional update, i. e., evaluating the particle's fitness after updating each variable of its position. Separate consideration of the impact of each variable decreases the occurrence of incomparable relations, thus improves the performance. Experimental results validate the efficiency of our algorithm.

The amount of information available on the Web has grown considerably in recent years, leading to the need to structure it in order to access it in a quick and accurate way. In order to develop techniques to automate the structuring process, the Knowledge Base Population (KBP) track of the Text Analysis Conference (TAC) was created. This forum aims to encourage research in automated systems capable of capturing knowledge in unstructured information. One of the tasks proposed in the context of the KBP track is named entity linking, and its goal is to link named entities mentioned in a document to instances in a reference knowledge base built from Wikipedia. This paper focuses on the entity linking task in the context of KBP 2010, where two different varieties of this task were considered, depending on whether the use of the text from Wikipedia was allowed or not. Specifically, the paper proposes a set of modifications to a system that participated in KBP 2010, named WikiIdRank, in order to improve its performance. The different modifications were evaluated in the official KBP 2010 corpus, showing that the best combination increases the accuracy of the initial system in a 7.04%. Though the resultant system, named WikiIdRank++, is unsupervised and does not take advantage of Wikipedia text, a comparison with other approaches in KBP indicates that the system would rank as 4th (out of 16) in the global comparison, outperforming other approaches that use human supervision and take advantage of Wikipedia textual contents. Furthermore, the system would rank as 1st in the category of systems that do not use Wikipedia text.

This paper is about designing, developing and training a neural network for short-term forecasting of buy-sell trends in foreign exchange markets. We use a set of established financial technical indicators as inputs to the neural network and we develop the architecture to predict a trend and then train the network based on the accuracy of the prediction. We perform extensive real time testing with the closing prices (one per minute) of the USD/EUR exchange rates for a one-year period. The overall approach delivers a system that predicts trends substantially better than individual technical indicators.

Clustering has been applied in a wide variety of disciplines and has also been utilized in many scientific areas. Usually, clustering algorithms construct either clusters of rows or clusters of columns of the input data matrix. Biclustering is a methodology where biclusters are formed by both a subset of rows and a subset of columns, such that objects represented by the first are the most similar to each other when compared over the latter. In this paper, we introduce a new biclustering technique, based on association rule mining, which can support different well-known biclustering models proposed in the literature. Experimental tests demonstrate the accuracy and efficiency of the proposed technique with respect to well known related ones.