Search

Choosing which fast Nearest Neighbour search algorithm to use depends on the task we face. Usually kd-tree search algorithm is selected when the similarity function is the Euclidean or the Manhattan distances. Generic fast search algorithms (algorithms that works with any distance function) are only used when there is not specific fast search algorithms for the involved distance function. In this work we show that in real data problems generic search algorithms (i.e. MDF-tree) can be faster that specific ones (i.e. kd-tree).

Many fast similarity search techniques relies on the use of pivots (specially selected points in the data set). Using these points, specific structures (indexes) are built speeding up the search when queering. Usually, pivot selection techniques are incremental, being the first one randomly chosen. This article explores several techniques to choose the first pivot in a tree-based fast similarity search technique. We provide experimental results showing that an adequate choice of this pivot leads to significant reductions in distance computations and time complexity. Moreover, most pivot tree-based indexes emphasizes in building balanced trees. We provide experimentally and theoretical support that very unbalanced trees can be a better choice than balanced ones.

Interactive Pattern Recognition concepts and techniques are applied to problems with structured output; i.e., problems in which the result is not just a simple class label, but a suitable structure of labels. For illustration purposes (a simplification of) the problem of Human Karyotyping is considered. Results show that a) taking into account label dependencies in a karyogram significantly reduces the classical (non-interactive) chromosome label prediction error rate and b) they are further improved when interactive processing is adopted.

Similarity search is a widely employed technique in Pattern Recognition. In order to speed up the search many indexing techniques have been proposed. However, the majority of the proposed techniques are static, that is, a fixed training set is used to build up the index. This characteristic becomes a major problem when these techniques are used in dynamic (interactive) systems. In these systems updating the training set is necessary to improve its performance. In this work, we explore the surprising efficiency of a naïve algorithm that allows making incremental insertion in a previously known index: the MDF-tree.

Multiplicity Automata are devices that implement functions from a string space to a field. Usually the real number’s field is used. From a learning point of view there exists some algorithms that are able to identify any multiplicity automaton from membership and equivalence queries. In this work we realize that those algorithms can also be used if the algebraic structure of a field is relaxed to a divisive ring structure, that is, the commutativity of the product operation is dropped. Moreover, we define an algebraic structure, which is an extension of the string monoid, that allows the identification of any transduction that can be realized by finite state machines without empty-transitions.

Nearest neighbour search is a simple technique widely used in Pattern Recognition tasks. When the dataset is large and/or the dissimilarity computation is very time consuming the brute force approach is not practical. In such cases, some properties of the dissimilarity measure can be exploited in order to speed up the search. In particular, the metric properties of some dissimilarity measures have been used extensively in fast nearest neighbour search algorithms to avoid dissimilarity computations. Recently, a distance table based pruning rule to reduce the average number of distance computations in hierarchical search algorithms was proposed. In this work we show the effectiveness of this rule compared to other state of the art algorithms. Moreover, we propose some guidelines to reduce the space complexity of the rule.

Due to its robustness to outliers, many Pattern Recognition algorithms use the median as a representative of a set of points. A special case arises in Syntactical Pattern Recognition when the points (prototypes) are represented by strings. However, when the edit distance is used, finding the median becomes a NP-Hard problem. Then, either the search is restricted to strings in the data (set-median) or some heuristic approach is applied. In this work we use the (conditional) stochastic edit distance instead of the plain edit distance. It is not yet known if in this case the problem is also NP-Hard so an approximation algorithm is described. The algorithm is based on the extension of the string structure to multistrings (strings of stochastic vectors where each element represents the probability of each symbol) to allow the use of the Expectation Maximization technique. We carry out some experiments over a chromosomes corpus to check the efficiency of the algorithm.

In this paper, we present SEDiL, a Software for Edit Distance Learning. SEDiL is an innovative prototype implementation grouping together most of the state of the art methods that aim to automatically learn the parameters of string and tree edit distances.

Some fast nearest neighbor search (NNS) algorithms using metric properties have appeared in the last years for reducing computational cost. Depending on the structure used to store the training set, different strategies to speed up the search have been defined. For instance, pruning rules avoid the search of some branches of a tree in a tree-based search algorithm. In this paper, we propose a new and simple pruning rule that can be used in most of the tree-based search algorithms. All the information needed by the rule can be stored in a table (at preprocessing time). Moreover, the rule can be computed in constant time. This approach is evaluated through real and artificial data experiments. In order to test its performance, the rule is compared to and combined with other previously defined rules.

Recently an algorithm to identify in the limit with polynomial time and data Left Deterministic Linear Languages (Left DLL) and, consequently Right DLL, was proposed. In this paper we show that the class of the Left-Right DLL formed by the union of both classes is also identifiable. To do that, we introduce the notion of n-negative characteristic sample, that is a sample that forces an inference algorithm to output an hypothesis of size bigger than n when strings from a non identifiable language are provided.