Search

Kernel methods provide a computational framework to integrate heterogeneous biological data from different sources for a wide range of learning algorithms by designing a kernel for each di®erent information source and combining them in a unique kernel through simple mathematical operations. We develop here a novel technique for weighting kernels based on their von Neumann entropy. This permits to assess the kernel quality without using label information, and to integrate kernels before the beginning of the learning process. Moreover, we carry out a comparison with the unweighted kernel summation and a popular technique based on semi-definite programming on kernel integration benchmark data sets. Finally, we empirically study the variation of the performance of a support vector machine classi¯er considering pairs of kernels combined in di®erent ratios, and we show how, surprisingly, the unweighted sum of kernels seems to lead to the same performance than a more complex weighting schema.

People belong to different communities: business communities, Web 2.0 communities, just to name a few. In this chapter we show that experience acquired by people in communities constitute community culture. We introduce the problem of culture transfer between or within communities and propose a domain-independent approach for transferring community culture. First, we formalize the notion of culture, which includes behavior, knowledge, artifacts, best practices, etc. Second, using this formalism, we propose the Implicit Culture Framework, which is an agent-based framework for transferring behavior between community members or between communities. Finally, we present and evaluate a system for web service discovery developed using the Implicit Culture Framework.

Local SVM is a classification method that combines instance-based learning and statistical machine learning. It builds an SVM on the feature space neighborhood of the query point in the training set and uses it to predict its class. There is both empirical and theoretical evidence that Local SVM can improve over SVM and kNN in terms of classification accuracy, but the computational cost of the method permits the application only on small datasets. Here we propose FastLSVM, a classifier based on Local SVM that decreases the number of SVMs that must be built in order to be suitable for large datasets. FastLSVM precomputes a set of local SVMs in the training set and assigns to each model all the points lying in the central neighborhood of the k points on which it is trained. The prediction is performed applying to the query point the model corresponding to its nearest neighbor in the training set. The empirical evaluation we provide points out that FastLSVM is a good approximation of Local SVM and its computational performances on big datasets (a large artificial problem with 100000 samples and on a very large real problem with more than 500000 samples) dramatically ameliorate SVM ones improving also the generalization accuracies.

Globalization makes culture no more bound to a geographical area, race or religion. Multi-national companies, software developers, scientists need to take into account cultural differences when delivering products to people. The first step in dealing with culture consists in defining and representing culture of the targeted community. AI literature addressed issues of sociality, collaboration, and coordination in agent societies, but did not target the problem of defining and representing culture of a community. In this paper, we propose a formal definition of culture of a set of agents. It generalizes existing definitions of culture and it is operational in the sense that it can be applied for characterization and comparison of culture(s) existing in various communities.

With the advent of Web 2.0 one can speak about best practices and specific knowledge existing in a community of users of an application. This suggest that there is a kind of “culture” growing around such applications, and this culture is created by the communities. Moreover, cultures of communities “interact” leading to changes in applications, i.e. introducing tabbed browsing in NetCaptor browser in 1998 lead to its appearance in Opera in 2000, Firefox in 2001, and IE in 2006. In this paper we propose a model of culture in communities of Web 2.0 applications and other social software. Such model can be applied for characterization and comparison of culture(s) existing in communities for better understanding the use of systems by people. We provide an operative definition of culture of a community and illustrate by examples how our model of culture can be applied in different contexts in systems for communities in Web 2.0 and in software in general.

In the field of statistical machine learning, the integration of kernel methods with local information has been proposed through locality-improved kernels for Support Vector Machines (SVM) that make use of prior information, local kernels and local SVM that apply the SVM approach only on the subset of points close to the testing one. Here we propose a novel family of operators on kernels able to integrate the local information into any kernel without prior information obtaining quasi-local kernels. The quasi-local kernels maintain the possibly global properties of the input kernel and they increase the kernel value as the points get closer in the feature space of the input kernel. The operators combine the input kernel with a locality-dependent term, and accept two parameters that regulate the width of the exponential influence of points in the locality-dependent term and the balancing between the two terms. Experiments carried out with data-dependent systematic selection of the parameters of the operators (i.e. without the need for model selection phase on the obtained kernels) on a total of 33 datasets with different characteristics and application domains, achieve very good results.

To some extent the problem of noise reduction in machine learning has been finessed by the development of learning techniques that are noise-tolerant. However, it is difficult to make instance-based learning noise tolerant and noise reduction still plays an important role in k-nearest neighbour classification. There are also other motivations for noise reduction, for instance the elimination of noise may result in simpler models or data cleansing may be an end in itself. In this paper we present a novel approach to noise reduction based on local Support Vector Machines (LSVM) which brings the benefits of maximal margin classifiers to bear on noise reduction. This provides a more robust alternative to the majority rule on which almost all the existing noise reduction techniques are based. Roughly speaking, for each training sample an SVM is trained on its neighbourhood and if the SVM classification for the central sample disagrees with its actual class there is evidence in favour of removing it from the training set. We provide an empirical evaluation on 15 real datasets showing improved classification accuracy when using training data edited with our method as well as specific experiments regarding the spam filtering application domain. We present a further evaluation on two artificial datasets where we analyse two different types of noise (Gaussian sample noise and mislabelling noise) and the influence of different class densities. The conclusion is that LSVM noise reduction is significatively better than the other analysed algorithms for real datasets and for artificial datasets perturbed by Gaussian noise and in presence of uneven class densities.

Multisite phosphorylation is known to be an important and dynamic mechanism for regulating the activity of transcription factors. Here we propose a stochastic pi-calculus modelling approach able to handle the complexity of post-translational modifications and to overcome the limitations of the ordinary differential equations based methods. The model can be applied without a priori assumptions to every (multisite) phosphorylation regulation for which some basic rates are known or can be indirectly set with experimental data. We apply it to the multisite phosphorylation of the Pho4 transcription factor that plays a crucial role in the phosphate starvation signalling in Saccharomyces cerevisiae using available in vitro experiments for the model tuning and validation. The innovative modelling of the sub-path with the stochastic pi-calculus allows quantitative analyses of the kinetic characteristics of the Pho4 phosphorylation, the different phosphorylation dynamics for each site (possibly combined) and the variation of the kinase activity as the reaction goes to completion. One of the performed predictions indicates that the Pho80-Pho85 kinase activity on the Pho4 substrate is nearly distributive and not semi-processive as previously found analysing only the phosphoform concentrations in vitro. This result is obtained because the model can consider and quantify the binding events without phosphorylations that cannot be experimentally measured. Thanks to the compositionality property of process algebras, we also developed the whole PHO pathway model that gives new suggestions and confirmations about its general behaviour. The potentialities of process calculi based in silico simulations for biological systems are highlighted and discussed.

Wang in a PAMI paper proposed Neighborhood Counting Measure (NCM) as a similarity measure for the k-nearest neighbors classification algorithm. In his paper, Wang mentioned Minimum Risk Metric (MRM) an earlier method based on the minimization of the risk of misclassification. However, Wang did not compare NCM with MRM because of its allegedly excessive computational load. In this letter, we empirically compare NCM against MRM on k-NN with k=1, 3, 5, 7 and 11 with decision taken with a voting scheme and k=21 with decision taken with a weighted voting scheme on the same datasets used by Wang. Our results shows that MRM outperforms NCM for most of the k values tested. Moreover, we show that the MRM computation is not so probihibitive as indicated by Wang. ©2009 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.

The complexity of the selection procedure of a genetic algorithm that requires reordering, if we restrict the class of the possible fitness functions to varying fitness functions, is O(N logN) where N is the size of the population. The Quantum Genetic Optimization Algorithm (QGOA) exploits the power of quantum computation in order to speed up genetic procedures. While the quantum and classical genetic algorithms use the same number of generations, the QGOA outperforms the classical one in identifying the high-fitness subpopulation at each generation. In QGOA the classical fitness evaluation and selection procedures are replaced by a single quantum procedure. We show that the complexity of our QGOA is O(1) in terms of number of oracle calls in the selection procedure. Such theoretical results are confirmed by the simulations of the algorithm.

The retrieval of cases plays important role in CBR systems. Current solutions to the problem of case retrieval adopt various similarity metrics such as utility-based, adaptation-guided similarity, and collaborative retrieval. In this paper, we present a collaborative architecture for case retrieval, where users share their experience in using cases and interact with one another by means of software agents. Past experiences are used to improve case retrieval. We describe a system for choosing software patterns as an implementation of the proposed architecture. Experimental results show that the system performs better than conventional retrieval engines.

The paradigmatic shift occurred in biology that led first to high-throughput experimental techniques and later to computational systems biology must be applied also to the analysis paradigm of the relation between local models and data to obtain an effective prediction framework. In this work we show that the new relation between systems biology models and high-throughput data permits new integrations on the systemic scale like the use of in silico predictions to support the mining of gene expression datasets. We introduce a unifying notational framework in which we propose two applications concerning the use of system level models to support the differential analysis of microarray expression data. The approach is tested with a specific microarray experiment on the phosphate system in Saccharomyces cerevisiae and a computational model of the PHO pathway that supports the systems biology concepts.

In this paper we evaluate the performance of the highest probability SVM nearest neighbor (HP-SVM-NN) classifier, which combines the ideas of the SVM and k-NN classifiers, on the task of spam filtering, using the pure SVM classifier as a quality baseline. To classify a sample the HP-SVM-NN classifier does the following: for each k in a predefined set {k1, ..., kN} it trains an SVM model on k nearest labeled samples, uses this model to classify the given sample, and transforms the output of SVM into posterior probabilities of the two classes using sigmoid approximation; than it selects that of the 2×N resulting answers which has the highest probability. The experimental evaluation shows, that in terms of ROC curves the algorithm is able to achieve higher accuracy than the pure SVM classifier.

Email spam is one of the major problems of the today’s Internet, bringing financial damage to companies and annoying individual users. Among the approaches developed to stop spam, filtering is an important and popular one. In this paper we give an overview of the state of the art of machine learning applications for spam filtering, and of the ways of evaluation and comparison of different filtering methods. We also provide a brief description of other branches of anti-spam protection and discuss the use of various approaches in commercial and noncommercial anti-spam software solutions.

Recommendation systems have proven to be useful in various application domains. However, current solutions are usually ad-hoc systems which are tightly-coupled with the application domain. We present the IC-Service, a recommendation service that can be included in any system in a loosely coupled way. The implementation follows the principles of service oriented computing and provides a solution to various problems arising in recommendation systems, e.g. to the problem of meta-recommendation systems development. Moreover, when properly configured, the IC-Service can be used by different applications (clients), and several independent instances of the IC-Service can collaborate to produce better recommendations. Service architecture and communication protocols are presented. The paper describes also ongoing work and applications based on the IC-Service.

Software patterns enable an efficient transfer of design experience by documenting common solutions to recurring design problems. They contain valuable knowledge that can be reused by others, in particular, by less experienced developers. Patterns have been published for system architecture and detailed design, as well as for specific application domains (e.g. agents and security). However, given the steadily growing number of patterns in the literature and online repositories, it can be hard for non-experts to select patterns appropriate to their needs, or even to be aware of the existing patterns. In this paper, we present a multi-agent system that supports developers in choosing patterns that are suitable for a given design problem. The system implements an implicit culture approach for recommending patterns to developers based on the history of decisions made by other developers regarding which patterns to use in related design problems. The recommendations are complemented with the documents from a pattern repository that can be accessed by the agents. The paper includes a set of experimental results obtained using a repository of security patterns. The results prove the viability of the proposed approach.

Very often people tend to behave as other people behaved previously. This happens in many different situations, for instance when one has to choose the path in a forest or when she/he has to select a link in the web. Social Navigation aims at providing assistance in such situations supporting the decision making process. Implicit Culture is a recent approach in which people are encouraged (induced) to behave according to usual behavior of the community. This paper shows similarities between Social Navigation and Implicit Culture and it presents a case study about user preferences learning.

The number of accessible web pages in Internet increases every day and it becomes very difficult to deal with such a huge source of information. There exist several approaches aimed to provide user with high-quality links extracted from the thousands of irrelevant ones. We present a system Implicit that combines recommender system and multi-agent system approaches and is intended to be used within community of people with similar interests. It produces suggestions by using implicit knowledge of the members of a community and complementing these suggestions with results from search engines. Agents within the system interact one another and share knowledge in order to increase quality of the recommendations by using similarities in the behavior of different users.

It is very difficult for beginners to define and find the most relevant literature in a research field. They can search on the web or look at the most important journals and conference proceedings, but it should be much better to receive suggestions directly from experts of the field. Unfortunately, this is not always possible and systems like CiteSeer and GoogleScholar become extremely useful for beginners (and not only). In this paper, we present an agent-based system that facilitates scientific publications search. Users interacting with their personal agents produce a transfer of knowledge about relevant publications from experts to beginners. Each personal agent observes how publications are used and induces behavioral patterns that are used to create more effective recommendations. Feedback exchange allows agents to share their knowledge and virtual communities of cloned experts can be created to support novice users. We present also a set of experimental results obtained using CiteSeer as a source of information, that show the effectiveness of our approach.

Design patterns represent solutions to problems that have been proved to be useful in different contexts and can be reused. An experienced programmer can choose a suitable pattern for a given problem effectively. However, for an inexperienced programmer this is a very hard task. We propose a multi-agent system that supports programmers in choosing the design pattern suitable for the given problem. Personal agents in our system produce knowledge transfer among users, allowing for the reuse of experience in choosing design patterns.

The complexity of the selection procedure of a genetic algorithm that requires reordering, if we restrict the class of the possible fitness functions to non–local or time–dependent fitness functions, is O(N logN) where N is the size of the population. Quantum Genetic Algorithm(QGA) exploits the power of quantum computation in order to speed up genetic procedures. In QGA the classical fitness evaluation and selection procedures are replaced by a single quantum procedure. QGA outperforms a typical classical genetic algorithm. We show that the complexity of our QGA is O(1) in terms of number of oracle calls in the selection procedure. Such theoretical results are confirmed by the simulations of the algorithm.

The paper presents an empirical research whose goal is to study the interaction between a Case-Based Reasoning (CBR) system and users without knowledge of CBR. This interaction, that here is called CBR naïve interaction, can occur frequently in web-based CBR systems and consequently in CBR-based e-commerce systems. The research was led during the usability assessment of COOL-TOUR a web-based CBR system for tourist culture support. The paper describes methods and results of the usability tests as well as a preliminary experiment. The quantitative results are not final but qualitative evidence suggests that the user adopts different models of interaction.

Medical applications usually used Radial Basis Function Networks just as Artificial Neural Networks. However, RBFNs are Knowledge-Based Networks that can be interpreted in several way: Artificial Neural Networks, Regularization Networks, Support Vector Machines, Wavelet Networks, Fuzzy Controllers, Kernel Estimators, Instanced-Based Learners. A survey of their interpretations and of their corresponding learning algorithms is provided as well as a brief survey on dynamic learning algorithms. RBFNs' interpretations can suggest applications that are particularly interesting in medical domains.

The goal of our research is to provide techniques that can assess and validate the results of SVM-based analysis of microarray data. We present preliminary results of the effect of mislabeled training samples. We conducted several systematic experiments on artificial and real medical data using SVMs. We systematically flipped the labels of a fraction of the training data. We show that a relatively small number of mislabeled examples can dramatically decrease the performance as visualized on the ROC graphs. This phenomenon persists even if the dimensionality of the input space is drastically decreased, by using for example feature selection. Moreover we show that for SVM recursive feature elimination, even a small fraction of mislabeled samples can completely change the resulting set of genes. This work is an extended version of the previous paper [MBN04].

We present an implementation of a multi-agent system that aims at solving the problem of tacit knowledge transfer by means of experiences sharing. In particular, we consider experiences of use of pieces of information. Each agent incorporates a system for implicit culture support (SICS) whose goal is to realize the acceptance of the suggested information. The SICS permits a transparent (implicit) sharing of the information about the use, e.g., requesting and accepting pieces of information.

The amount of information on Internet is increasing very fast and, as a result, search becomes more and more a harder task. A common solution is to use authority-based search engines. However, for a community of people with similar interests, quality of results can be improved exploiting also implicit knowledge. We propose an agent-based recommendation system for supporting communities of people in searching the web by means of a popular search engine. Agents use data mining techniques in order to learn and discover users behaviors, and they interact to share knowledge about the users. We also present set of experimental results showing in terms of precision and recall how interaction increases the performance of the system.

This paper discusses the uses of context in knowledge representation and reasoning (KRR). We propose to partition the theories of context brought forward in KRR into two main classes, which we call divide-and-conquer and compose-and-conquer. We argue that this partition provides a possible explanation of why in KRR context is used to solve different types of problems, or to address the same problems from very different perspectives. The problems we use to illustrate this point are the problem of generality, the formalization of propositional attitudes, knowledge and data integration.

We present an implementation of a multi-agent system that aims at solving the problem of tacit knowledge transfer by means of experiences sharing. In particular, we consider experiences of use of pieces of information. Each agent incorporates a system for implicit culture support (SICS) whose goal is to realize the acceptance of the suggested information. The SICS permits a transparent (implicit) sharing of the information about the use, e.g., requesting and accepting pieces of information.

Globalization makes culture no more bound to a geographical area, race or religion. Multi-national companies, software developers, scientists need to take into account cultural differences when delivering products to people. The first step in dealing with culture consists in defining and representing culture of the targeted community. AI literature addressed issues of sociality, collaboration, and coordination in agent societies, but did not target the problem of defining and representing culture of a community. In this paper, we propose a formal definition of culture of a set of agents. It generalizes existing definitions of culture and it is operational in the sense that it can be applied for characterization and comparison of culture(s) existing in various communities.

Wang in a PAMI paper proposed Neighborhood Counting Measure (NCM) as a similarity measure for the k-nearest neighbors classification algorithm. In his paper, Wang mentioned Minimum Risk Metric (MRM) an earlier method based on the minimization of the risk of misclassification. However, Wang did not compare NCM with MRM because of its allegedly excessive computational load. In this letter, we empirically compare NCM against MRM on k-NN with k=1, 3, 5, 7 and 11 with decision taken with a voting scheme and k=21 with decision taken with a weighted voting scheme on the same datasets used by Wang. Our results shows that MRM outperforms NCM for most of the k values tested. Moreover, we show that the MRM computation is not so probihibitive as indicated by Wang. ©2009 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.

With the advent of Web 2.0 one can speak about best practices and specific knowledge existing in a community of users of an application. This suggest that there is a kind of “culture” growing around such applications, and this culture is created by the communities. Moreover, cultures of communities “interact” leading to changes in applications, i.e. introducing tabbed browsing in NetCaptor browser in 1998 lead to its appearance in Opera in 2000, Firefox in 2001, and IE in 2006. In this paper we propose a model of culture in communities of Web 2.0 applications and other social software. Such model can be applied for characterization and comparison of culture(s) existing in communities for better understanding the use of systems by people. We provide an operative definition of culture of a community and illustrate by examples how our model of culture can be applied in different contexts in systems for communities in Web 2.0 and in software in general.

The combination of maximal margin classifiers and k-nearest neighbors rule constructing an SVM on the neighborhood of the test sample in the feature space (called kNNSVM), was presented as a promising way of improving classification accuracy. Since no extensive validation of the method was performed yet, in this work we test the kNNSVM method on 13 widely used datasets using four different kernels obtaining good classification results. Moreover we present two artificial datasets in which kNNSVM performs substantially better than SVM with RBF kernel. Statistically significant testing of the method as well as the results on the artificial datasets, lead us to conclude that kNNSVM performs sensibly better than SVM.

Multisite phosphorylation is known to be an important and dynamic mechanism for regulating the activity of transcription factors. Here we propose a stochastic pi-calculus modelling approach able to handle the complexity of post-translational modifications and to overcome the limitations of the ordinary differential equations based methods. The model can be applied without a priori assumptions to every (multisite) phosphorylation regulation for which some basic rates are known or can be indirectly set with experimental data. We apply it to the multisite phosphorylation of the Pho4 transcription factor that plays a crucial role in the phosphate starvation signalling in Saccharomyces cerevisiae using available in vitro experiments for the model tuning and validation. The innovative modelling of the sub-path with the stochastic pi-calculus allows quantitative analyses of the kinetic characteristics of the Pho4 phosphorylation, the different phosphorylation dynamics for each site (possibly combined) and the variation of the kinase activity as the reaction goes to completion. One of the performed predictions indicates that the Pho80-Pho85 kinase activity on the Pho4 substrate is nearly distributive and not semi-processive as previously found analysing only the phosphoform concentrations in vitro. This result is obtained because the model can consider and quantify the binding events without phosphorylations that cannot be experimentally measured. Thanks to the compositionality property of process algebras, we also developed the whole PHO pathway model that gives new suggestions and confirmations about its general behaviour. The potentialities of process calculi based in silico simulations for biological systems are highlighted and discussed.

The paradigmatic shift occurred in biology that led first to high-throughput experimental techniques and later to computational systems biology must be applied also to the analysis paradigm of the relation between local models and data to obtain an effective prediction framework. In this work we show that the new relation between systems biology models and high-throughput data permits new integrations on the systemic scale like the use of in silico predictions to support the mining of gene expression datasets. We introduce a unifying notational framework in which we propose two applications concerning the use of system level models to support the differential analysis of microarray expression data. The approach is tested with a specific microarray experiment on the phosphate system in Saccharomyces cerevisiae and a computational model of the PHO pathway that supports the systems biology concepts. This is the preliminary version of a paper that was published in Journal of Integrative Bioinformatics, 5(1):87-103, 2008. The original publication is available at http://journal.imbio.de

It is very difficult for beginners to define and find the most relevant literature in a research field. They can search on the web or look at the most important journals and conference proceedings, but it should be much better to receive suggestions directly from experts of the field. Unfortunately, this is not always possible and systems like CiteSeer and GoogleScholar become extremely useful for beginners (and not only). In this paper, we present an agent-based system that facilitates scientific publications search. Users interacting with their personal agents produce a transfer of knowledge about relevant publications from experts to beginners. Each personal agent observes how publications are used and induces behavioral patterns that are used to create more effective recommendations. Feedback exchange allows agents to share their knowledge and virtual communities of cloned experts can be created to support novice users. We present also a set of experimental results obtained using CiteSeer as a source of information, that show the effectiveness of our approach.

Medical applications usually used Radial Basis Function Networks just as Artificial Neural Networks. However, RBFNs are Knowledge-Based Networks that can be interpreted in several way: Artificial Neural Networks, Regularization Networks, Support Vector Machines, Wavelet Networks, Fuzzy Controllers, Kernel Estimators, Instanced-Based Learners. A survey of their interpretations and of their corresponding learning algorithms is provided as well as a brief survey on dynamic learning algorithms. RBFNs' interpretations can suggest applications that are particularly interesting in medical domains.

In the field of statistical machine learning, the integration of kernel methods with local information has been proposed through locality-improved kernels for Support Vector Machines (SVM) that make use of prior information, local kernels and local SVM that apply the SVM approach only on the subset of points close to the testing one. Here we propose a novel family of operators on kernels able to integrate the local information into any kernel without prior information obtaining quasi-local kernels. The quasi-local kernels maintain the possibly global properties of the input kernel and they increase the kernel value as the points get closer in the feature space of the input kernel. The operators combine the input kernel with a locality-dependent term, and accept two parameters that regulate the width of the exponential influence of points in the locality-dependent term and the balancing between the two terms. Experiments carried out with data-dependent systematic selection of the parameters of the operators (i.e. without the need for model selection phase on the obtained kernels) on a total of 33 datasets with different characteristics and application domains, achieve very good results.

We present an implementation of a multi-agent system that aims at solving the problem of tacit knowledge transfer by means of experiences sharing. In particular, we consider experiences of use of pieces of information. Each agent incorporates a system for implicit culture support (SICS) whose goal is to realize the acceptance of the suggested information. The SICS permits a transparent (implicit) sharing of the information about the use, e.g., requesting and accepting pieces of information.

Globalization makes culture no more bound to a geographical area, race or religion as was previously considered in anthropology. With the advent of Web 2.0 it becomes appropriate to speak about the culture of online communities in general, without specific ties to country or nation. Most current approaches to the development of systems for online communities implement some pre-defined culture rather than supporting the actual culture of the community for which the system is being developed. This technical report presents a formal definition of culture of a set of agents. Our definition covers many previous definitions existing in anthropology, and it can be used to model the culture of online communities. This is the first version of the formalism and it does not introduce states. However, representing a snapshot of the culture in a certain moment is the first step towards a more complex formalism that includes states (the work on the latter is ongoing).

A computationally efficient approach to local learning with kernel methods is presented. The Fast Local Kernel Support Vector Machine (FaLK-SVM) trains a set of local SVMs on redundant neighbourhoods in the training set and an appropriate model for each query point is selected at testing time according to a proximity strategy. Supported by a recent result by Zakai and Ritov (2009) relating consistency and localizability, our approach guarantees high generalization ability by dividing the separation function in local optimization problems that can be handled very efficiently. The introduction of a fast local model selection further speeds-up the learning process. Learning and complexity bounds are derived for FaLK-SVM, and the empirical evaluation of the approach (with datasets up to 3 million points) showed that it is much faster and more accurate and scalable than state-of-the-art accurate and approximated SVM solvers at least for non high-dimensional datasets. More generally, we show that locality can be an important factor to sensibly speed-up learning approaches and kernel methods, differently from other recent techniques that tend to dismiss local information in order to improve scalability.

We present an implementation of a multi-agent system that aims at solving the problem of tacit knowledge transfer by means of experiences sharing. In particular, we consider experiences of use of pieces of information. Each agent incorporates a system for implicit culture support (SICS) whose goal is to realize the acceptance of the suggested information. The SICS permits a transparent (implicit) sharing of the information about the use, e.g., requesting and accepting pieces of information.

The retrieval of cases plays important role in CBR systems. Current solutions to the problem of case retrieval adopt various similarity metrics such as utility-based, adaptation-guided similarity, and collaborative retrieval. In this paper, we present a collaborative architecture for case retrieval, where users share their experience in using cases and interact with one another by means of software agents. Past experiences are used to improve case retrieval. We describe a system for choosing software patterns as an implementation of the proposed architecture. Experimental results show that the system performs better than conventional retrieval engines.

In this paper we evaluate the performance of the highest probability SVM nearest neighbor (HP-SVM-NN) classifier, which combines the ideas of the SVM and k-NN classifiers, on the task of spam filtering, using the pure SVM classifier as a quality baseline. To classify a sample the HP-SVM-NN classifier does the following: for each k in a predefined set {k1, ..., kN} it trains an SVM model on k nearest labeled samples, uses this model to classify the given sample, and transforms the output of SVM into posterior probabilities of the two classes using sigmoid approximation; than it selects that of the 2×N resulting answers which has the highest probability. The experimental evaluation shows, that in terms of ROC curves the algorithm is able to achieve higher accuracy than the pure SVM classifier.

This paper discusses the uses of context in knowledge representation and reasoning (KRR). We propose to partition the theories of context brought forward in KRR into two main classes, which we call divide-and-conquer and compose-and-conquer. We argue that this partition provides a possible explanation of why in KRR context is used to solve different types of problems, or to address the same problems from very different perspectives. The problems we use to illustrate this point are the problem of generality, the formalization of propositional attitudes, knowledge and data integration.