Search

Heterogeneous systems employing CPUs and GPUs are becoming increasingly popular in large-scale data centers and cloud environments. In these platforms, sharing a GPU across different applications is an important feature to improve hardware utilization and system throughput. However, under scenarios where GPUs are competitively shared, some challenges arise. The decision on the simultaneous execution of different kernels is made by the hardware and depends on the kernels resource requirements. Besides that, it is very difficult to understand all the hardware variables involved in the simultaneous execution decisions in order to describe a formal allocation method. In this work, we use machine learning techniques to understand how the resource requirements of the kernels from the most important GPU benchmarks impact their concurrent execution. We focus on making the machine learning algorithms capture the hidden patterns that make a kernel interfere in the execution of another one when they are submitted to run at the same time. The techniques analyzed were k -NN, Logistic Regression, Multilayer Perceptron and XGBoost (which obtained the best results) over the GPU benchmark suites, Rodinia, Parboil and SHOC. Our results showed that, from the features selected in the analysis, the number of blocks per grid, number of threads per block, and number of registers are the resource consumption features that most affect the performance of the concurrent execution.

Knowledge bases, and their representations in the form of knowledge graphs (KGs), are naturally incomplete. Since scientific and industrial applications have extensively adopted them, there is a high demand for solutions that complete their information. Several recent works tackle this challenge by learning embeddings for entities and relations, then employing them to predict new relations among the entities. Despite their aggrandizement, most of those methods focus only on the local neighbors of a relation to learn the embeddings. As a result, they may fail to capture the KGs' context information by neglecting long-term dependencies and the propagation of entities' semantics. In this manuscript, we propose {\AE}MP (Attention-based Embeddings from Multiple Patterns), a novel model for learning contextualized representations by: (i) acquiring entities' context information through an attention-enhanced message-passing scheme, which captures the entities' local semantics while focusing on different aspects of their neighborhood; and (ii) capturing the semantic context, by leveraging the paths and their relationships between entities. Our empirical findings draw insights into how attention mechanisms can improve entities' context representation and how combining entities and semantic path contexts improves the general representation of entities and the relation predictions. Experimental results on several large and small knowledge graph benchmarks show that {\AE}MP either outperforms or competes with state-of-the-art relation prediction methods., Comment: Accepted to publication at Knowledge-based Systems, 2022

International audience; Training Deep Learning (DL) models require adjusting a series of hyperparameters. Although there are several tools to automatically choose the best hyperparameter configuration, the user is still the main actor to take the final decision. To decide whether the training should continue or try different configurations, the user needs to analyze online the hyperparameters most adequate to the training dataset, observing metrics such as accuracy and loss values. Provenance naturally represents data derivation relationships (i.e., transformations, parameter values, etc.), which provide important support in this data analysis. Most of the existing provenance solutions define their own and proprietary data representations to support DL users in choosing the best hyperparameter configuration, which makes data analysis and interoperability difficult. We present Keras-Prov and its extension, named Keras-Prov++, which provides an analytical dashboard to support online hyperparameter fine-tuning. Different from the current mainstream solutions, Keras-Prov automatically captures the provenance data of DL applications using the W3C PROV recommendation, allowing for hyperparameter online analysis to help the user deciding on changing hyperparameters' values after observing the performance of the models on a validation set. We provide an experimental evaluation of Keras-Prov++ using AlexNet and a real case study, named DenseED, that acts as a surrogate model for solving equations. During the online analysis, the users identify scenarios that suggest reducing the number of epochs to avoid unnecessary executions and fine-tuning the learning rate to improve the model accuracy.

Sentiment analysis in tweets is a research field of great importance, mainly due to the popularity of Twitter. However, collecting and annotating tweets is an expensive and time-consuming task, making that some domains have only a limited set of labeled data. A promising strategy to handle this issue is to leverage labeled domains rich in data to select instances that enrich target datasets. This paper proposes different strategies for selecting instances from a set of labeled source datasets in order to improve the performance of classifiers trained only with the target dataset. Different approaches are proposed, including similarity metrics and variations in the number of selected instances. The results show that the size of the training set plays an essential role in the predictive capacity of the classifier. Furthermore, the results point out the importance of taking into account diversity criteria when selecting the instances.

Handling relational data streams has become a crucial task, given the availability of pervasive sensors and Internet-produced content, such as social networks and knowledge graphs. In a relational environment, this is a particularly challenging task, since one cannot assure that the streams of examples are independent along the iterations. Thus, most relational learning systems are still designed to learn only from closed batches of data. Furthermore, in case there is a previously acquired model, these systems either would discard it or assuming it as correct. In this work, we propose an online relational learning algorithm that can handle continuous, open-ended streams of relational examples as they arrive. We employ techniques of theory revision to take advantage of the previously acquired model as a starting point, by finding where it should be modified to cope with the new examples, and automatically update it. We rely on the Hoeffding’s bound statistical theory to decide if the model must, in fact, be updated in accordance with the new examples. The proposed algorithm is built upon ProPPR statistical relational language, aiming at contemplating the uncertainty inherent to real data. Experimental results in social networks and entity co-reference datasets show the potential of the proposed approach compared to other relational learners.

Scientific Workflows (SWfs) have revolutionized how scientists in various domains of science conduct their experiments. The management of SWfs is performed by complex tools that provide support for workflow composition, monitoring, execution, capturing, and storage of the data generated during execution. In some cases, they also provide components to ease the visualization and analysis of the generated data. During the workflow’s composition phase, programs must be selected to perform the activities defined in the workflow specification. These programs often require additional parameters that serve to adjust the program’s behavior according to the experiment’s goals. Consequently, workflows commonly have many parameters to be manually configured, encompassing even more than one hundred in many cases. Wrongly parameters’ values choosing can lead to crash workflows executions or provide undesired results. As the execution of data- and compute-intensive workflows is commonly performed in a high-performance computing environment e.g., (a cluster, a supercomputer, or a public cloud), an unsuccessful execution configures a waste of time and resources. In this article, we present FReeP—Feature Recommender from Preferences, a parameter value recommendation method that is designed to suggest values for workflow parameters, taking into account past user preferences. FReeP is based on Machine Learning techniques, particularly in Preference Learning. FReeP is composed of three algorithms, where two of them aim at recommending the value for one parameter at a time, and the third makes recommendations for n parameters at once. The experimental results obtained with provenance data from two broadly used workflows showed FReeP usefulness in the recommendation of values for one parameter. Furthermore, the results indicate the potential of FReeP to recommend values for n parameters in scientific workflows.

Game agents are efficiently implemented through different AI techniques, such as neural network, reinforcement learning, and evolutionary search. Although there are many works for each approach, we present a critical analysis and comparison between them, suggesting a common benchmark and parameter configurations. The evolutionary strategy implements the NeuroEvolution of Augmenting Topologies algorithm, while the reinforcement learning agent leverages Q-Learning and Proximal Policy Optimization. We formulate and empirically compare this set of solutions using the Flappy Bird game as a test scenario. We also compare different representations of state and reward functions for each method. All methods were able to generate agents that can play the game, where the NEAT algorithm had the best results, reaching the goal of never losing.

An ontology formalises a number of dependent and related concepts in a domain, encapsulated as a terminology. Manually defining such terminologies is a complex, time-consuming and error-prone task. Thus, there is great interest for strategies to learn terminologies automatically. However, most of the existing approaches induce a single concept definition at a time, disregarding dependencies that may exist among the concepts. As a consequence, terminologies that are difficult to interpret may be induced. Thus, systems capable of learning all concepts within a single task, respecting their dependency, are essential for reaching concise and readable ontologies. In this paper, we tackle this issue presenting three terminology learning strategies that aim at finding dependencies among concepts, before, during or after they have been defined. Experimental results show the advantages of regarding the dependencies among the concepts to achieve readable and concise terminologies, compared to a system that learns a single concept at a time. Moreover, the three strategies are compared and analysed towards discussing the strong and weak points of each one.

Sentiment analysis on social media data can be a challenging task, among other reasons, because labeled data for training is not always available. Transfer learning approaches address this problem by leveraging a labeled source domain to obtain a model for a target domain that is different but related to the source domain. However, the question that arises is how to choose proper source data for training the target classifier, which can be made considering the similarity between source and target data using distance metrics. This article investigates the relation between these distance metrics and the classifiers’ performance. For this purpose, we propose to evaluate four metrics combined with distinct dataset representations. Computational experiments, conducted in the Twitter sentiment analysis scenario, showed that the cosine similarity metric combined with bag-of-words normalized with term frequency-inverse document frequency presented the best results in terms of predictive power, outperforming even the classifiers trained with the target dataset in many cases.

Arguably, player behavior profiling is one of the most relevant tasks of Game Analytics. However, to fulfill the needs of this task, gameplay data should be handled so that the player behavior can be profiled and even understood. Usually, gameplay data is stored as raw log-like files, from which gameplay metrics are computed. However, gameplay metrics have been commonly used as input to classify player behavior with two drawbacks: (1) gameplay metrics are mostly handcrafted and (2) they might not be adequate for fine-grain analysis as they are just computed after key events, such as stage or game completion. In this paper, we present a novel approach for player profiling based on provenance graphs, an alternative to log-like files that model causal relationships between entities in game. Our approach leverages recent advances in deep learning over graph representation of player states and its neighboring contexts, requiring no handcrafted features. We perform clustering on learned nodes representations to profile at a fine-grain the player behavior in provenance data collected from a multiplayer battle game and assess the obtained profiles through statistical analysis and data visualization.

Statistical machine learning algorithms usually assume that there is considerably-size data to train the models. However, they would fail in addressing domains where data is difficult or expensive to obtain. Transfer learning has emerged to address this problem of learning from scarce data by relying on a model learned in a source domain where data is easy to obtain to be a starting point for the target domain. On the other hand, real-world data contains objects and their relations, usually gathered from noisy environment. Finding patterns through such uncertain relational data has been the focus of the Statistical Relational Learning (SRL) area. Thus, to address domains with scarce, relational, and uncertain data, in this paper, we propose TreeBoostler, an algorithm that transfers the SRL state-of-the-art Boosted Relational Dependency Networks learned in a source domain to the target domain. TreeBoostler first finds a mapping between pairs of predicates to accommodate the additive trees into the target vocabulary. After, it employs two theory revision operators devised to handle incorrect relational regression trees aiming at improving the performance of the mapped trees. In the experiments presented in this paper, TreeBoostler has successfully transferred knowledge among several distinct domains. Moreover, it performs comparably or better than learning from scratch methods in terms of accuracy and outperforms a transfer learning approach in terms of accuracy and runtime.

Traditionally, algorithms that learn the structure of Bayesian Networks either start from an empty graph and add edges to it bit by bit or add/ remove/reverse edges in a randomly initialized graph. In both cases, the search space is constituted of all the nodes and edges connecting them. Searching within such a vast scope is a hard task, which gets worse in the presence of a dataset with missing values. However, it may be the case that an initial structure already exists and to make it reflect the set of examples it would be required to modify only a subset of the graph. Thus, instead of searching through the entire space of possible connections between the nodes, the problem could be reduced to selecting a subset of the edges and revising them. In this work, we present a novel algorithm for refining the structure of Bayesian networks from incomplete data, named BaBReN (Bayes Ball for Revising Networks). BaBReN has as ultimate goal to improve the inference value of the class variable. Thus, the algorithm tries to solve classification issues by proposing local modifications to the edges connecting the nodes that influence the erroneous classification. The Bayes Ball algorithm – based on the d-separation criteria – is responsible for selecting those relevant nodes. By focusing only on the influential nodes, BaBReN is executed independently of the number of variables in the domain. BaBReN is compared to a constraint-based algorithm (GS), a hybrid one (MMHC) and a score-based one (SEM with GHC), presenting better or competitive results regarding time and classification score.

The use of social media data to mine opinions during elections has emerged as an alternative to traditional election polls. However, relying on social media data in electoral scenarios comes with a number of challenges, such as tackling sentences with domain specific terms, texts full of hate speech, noisy, informal vocabulary, sarcasm and irony. Also, in Twitter, for instance, loss of context may occur due to the imposed limit of characters to the posts. Furthermore, prediction tasks that use machine learning require labeled datasets and it is not trivial to reliably annotate them during the short period of campaigns. Motivated by the aforementioned issues, we investigate if it is possible to use or mix curated datasets from other domains as a starting point to opinion mining tasks during elections. To avoid introducing a knowledge from the other domains that could end up by disturbing the task, we propose to use similarity metrics that point out whether or not the dataset should be used. In our approach, we conduct a case study using the 2018 Brazilian Presidential Elections and labeled datasets for sentiment analysis from other domains. To identify the similarity between the datasets, we use the Jaccard distance and a metric based on word embeddings. Our experimental results show that taking into account the (dis) similarity between different domains, it is possible to achieve results closer to the ones that would be achieved with classifiers trained with annotated datasets of the electoral domain.

Making artificial agents that learn how to play is a long-standing goal in the area of Game AI. Recently, several successful cases have emerged driven by Reinforcement Learning (RL) and neural network-based approaches. However, in most of the cases, the results have been achieved by training directly from pixel frames with valuable computational resources. In this paper, we devise agents that learn how to play the popular game of Bomberman by relying on state representations and RL-based algorithms without looking at the pixel level. To that, we designed five vector-based state representations and implemented Bomberman on the top of the Unity game engine through the ML-agents toolkit. We enhance the ML-agents algorithms by developing an Imitation-based learner (IL) that improves its model with the Actor-Critic Proximal-Policy Optimization (PPO) method. We compared this approach with a PPO-only learner that uses either a Multi-Layer Perceptron or a Long-Short Term-Memory network (LSTM). We conducted several pieces of training and tournament experiments by making the agents play against each other. The hybrid state representation and our IL followed by PPO learning algorithm achieve the best overall quantitative results, and we also observed that their agents learn a correct Bomberman behavior.

Smart cities initiatives have the potential to improve the life of citizens in a huge number of dimensions. One of them is the development of techniques and services capable of contributing to the enhancement of security public policies. Finding criminal patterns from historical data would arguably help in predicting and even preventing thefts and burglaries that continuously increase in urban centers worldwide. However, accessing such history and finding patterns across the interrelated crime occurrences data are challenging tasks, particularly to underdevelopment countries. In this paper, we address these problems by combining three techniques: we collect crime data from existing crowd-sourcing systems, we automatically induce patterns with relational machine learning, and we manage the entire process using scientific workflows. The framework developed under these lines is named CRiMINaL (Crime patteRn MachINe Learning). Experimental results conducted from a popular Brazilian source of data and a traditional relational learning system shows that CRiMINaL is a promising tool to induce interpretable models that can assist police departments on crime prevention.

Automated story generation is a desired feature in games and interactive media because it can control how a virtual world evolves so that it can be adapted to the players' choices. In order to have variety and quality in the generated stories, previous works have relied on simulation-based storytelling, in which a story is generated as their characters, represented as agents, try to achieve their goals. One challenge of this approach is to make the agents act more like human characters and less like omniscient intelligent beings. In this article, we present a perception model for simulation-based story generation that introduces errors into characters' knowledge, (mis)leading them to non-optimal, but still coherent, believable actions. The perception is executed using a description of the virtual world's elements using physical characteristics, and a pattern matching process that associates combinations of physical characteristics with predefined combinations of attributes, which are allowed to be wrong, and consequently may result in non-perfect interpretations of the world. We developed a story generation system from the proposed model and tested it with a version of the Little Red Riding Hood story, famous for its perception failure. Our results show interesting variations for the traditional known ending.

Deep Learning techniques have achieved impressive results over the last few years. However, they still have difficulty in producing understandable results that clearly show the embedded logic behind the inductive process. One step in this direction is the recent development of Neural Differentiable Programmers. In this paper, we designed a neural programmer that can be easily integrated into existing deep learning architectures, with similar amount of parameters to a single commonly used Recurrent Neural Network. Tests conducted with the proposal suggest that it has the potential to induce algorithms even without any kind of special optimization, achieving competitive results in problems handled by more complex RNN architectures.

Game Analytics comprises a set of techniques to analyze both the game quality and player behavior. To succeed in Game Analytics, it is essential to identify what is happening in a game (an effect) and track its causes. Thus, game provenance graph tools have been proposed to capture cause-and-effect relationships occurring in a gameplay session to assist the game design process. However, since game provenance data capture is guided by a set of strict predefined rules established by the game developers, the detection of long-range cause-and-effect relationships may demand huge coding efforts. In this paper, we contribute with a framework named PingUMiL that leverages the recently proposed graph embeddings to represent game provenance graphs in a latent space. The embeddings learned from the data pose as the features of a machine learning task tailored towards detecting long-range cause-and-effect relationships. We evaluate the generalization capacity of PingUMiL when learning from similar games and compare its performance to classical machine learning methods. The experiments conducted on two racing games show that (1) PingUMiL outperforms classical machine learning methods and (2) representation learning can be used to detect long-range cause-and-effect relationships in only partially observed game data provenance graphs.

Theory Revision from Examples is the process of repairing incorrect theories and/or improving incomplete theories from a set of examples. This process usually results in more accurate and comprehensible theories than purely inductive learning. However, so far, progress on the use of theory revision techniques has been limited by the large search space they yield. In this article, we argue that it is possible to reduce the search space of a theory revision system by introducing stochastic local search. More precisely, we introduce a number of stochastic local search components at the key steps of the revision process, and implement them on a state-of-the-art revision system that makes use of the most specific clause to constrain the search space. We show that with the use of these SLS techniques it is possible for the revision system to be executed in a feasible time, while still improving the initial theory and in a number of cases even reaching better accuracies than the deterministic revision process. Moreover, in some cases the revision process can be faster and still achieve better accuracies than an ILP system learning from an empty initial hypothesis or assuming an initial theory to be correct.

Deep learning models have recently achieved the state-of-the-art results on a well-known pedestrian detection dataset. However, such images were obtained from open scenarios with fixed imaging geometry parameters, which may produce a network not suitable for detecting a person in more general settings, such as the ones found in surveillance systems. As gathering and annotating data is a highly expensive manual task, we propose a methodology for artificially augmenting the positive training set with automatically generated local image affine and perspective transforms. Furthermore, to enrich the variability of background images, we include to the negative training set images that resemble human figures automatically obtained by the proposed methodology over images from commonly found surveillance scenarios. Extensive results show that by providing the enriched data as the input to a Convolutional Neural Network it is possible to precisely detect pedestrians in a number of public datasets. The data enrichment proposed here may also be used in other detectors based on supervised learning architectures, as the process is independent from the learning algorithm employed.

Stories have become an important element of games, since they can increase their immersion level by giving the players the context and the motivation to play. However, despite the interactive nature of games, their stories usually do not develop considering every decision and/or action the players are capable of, because depending on the game size, it would take too much effort to author alternative routes for all of them. To make these alternatives viable, an interesting solution would be to procedurally generate them, which could be achieved by using the story generation approaches already developed by many works of the storytelling field. Some of these approaches are based on the simulation of virtual worlds, in which the stories are generated by making the characters that inhabit the worlds act trying to reach their goals. The resulting actions and the world's reactions compose the final story. Since the actions are the building blocks of the stories, the characters' acting capabilities are determinant features of the generation potential of simulations. For instance, it is only possible to generate stories with deception if the characters are capable of deceiving each other. To allow the generation of stories where the characters are capable of manipulation, cooperation and other social behaviors by actively using what the others will do based on what they know and see, we propose a recursive planning approach that deals with the uncertainty of the others' knowledge and with a purposely error-prone perception simulation. To test our proposal we developed a story generation system and designed an adaptation of Little Red Riding Hood world as test scenario. With our approach, the system was capable of generating coherent story variations with deceptive actions.

Theory revision systems are designed to improve the accuracy of an initial theory, producing more accurate and comprehensible theories than purely inductive methods. Such systems search for points where examples are misclassified and modify them using revision operators. This includes trying to add antecedents to clauses usually following a top-down approach, considering all the literals of the knowledge base. Such an approach leads to a huge search space which dominates the cost of the revision process. ILP Mode Directed Inverse Entailment systems restrict the search for antecedents to the literals of the bottom clause. In this work the bottom clause and mode declarations are introduced in a first-order logic theory revision system aiming to improve the efficiency of the antecedent addition operation and, consequently, also of the whole revision process. Experimental results compared to revision system FORTE show that the revision process is on average 55 times faster, generating more comprehensible theories and still not significantly decreasing the accuracies obtained by the original revision process. Moreover, the results show that when the initial theory is approximately correct, it is more efficient to revise it than learn from scratch, obtaining significantly better accuracies. They also show that using the proposed theory revision system to induce theories from scratch is faster and generates more compact theories than when the theory is induced using a traditional ILP system, obtaining competitive accuracies.

Traditional machine learning algorithms require a dataset composed of homogeneous objects, randomly sampled from a single relation. However, real world tasks such as link prediction and entity resolution, require the representation of multiple relations, heterogeneous and structured data. Inductive Logic Programming (ILP) is a sub area of machine learning that induces structured hypotheses from multi-relational examples and background knowledge (BK) represented as logical clauses. With a few exceptions, most of the systems developed in ILP induce Horn-clauses and uses Prolog as their baseline inference engine. However, the recent development of efficient Answer Set Programming solvers points out that these can be a viable option to be the reasoning component of ILP systems, especially to address nonmonotonic reasoning. In this paper, we present dASBoT, a system that is capable of inducing extended normal rules mined from answer sets yielded from the examples and the BK. We show empirical evidence that dASBoT can support the task of relational identification by learning rules in three link prediction and two entity resolution tasks.

This book constitutes the thoroughly refereed post-proceedings of the 23rd International Conference on Inductive Logic Programming, ILP 2013, held in Rio de Janeiro, Brazil, in August 2013. The 9 revised extended papers were carefully reviewed and selected from 42 submissions. The conference now focuses on all aspects of learning in logic, multi-relational learning and data mining, statistical relational learning, graph and tree mining, relational reinforcement learning, and other forms of learning from structured data.

The game of chess has been a major testbed for research in artificial intelligence, since it requires focus on intelligent reasoning. Particularly, several challenges arise to machine learning systems when inducing a model describing legal moves of the chess, including the collection of the examples, the learning of a model correctly representing the official rules of the game, covering all the branches and restrictions of the correct moves, and the comprehensibility of such a model. Besides, the game of chess has inspired the creation of numerous variants, ranging from faster to more challenging or to regional versions of the game. The question arises if it is possible to take advantage of an initial classifier of chess as a starting point to obtain classifiers for the different variants. We approach this problem as an instance of theory revision from examples. The initial classifier of chess is inspired by a FOL theory approved by a chess expert and the examples are defined as sequences of moves within a game. Starting from a standard revision system, we argue that abduction and negation are also required to best address this problem. Experimental results show the effectiveness of our approach.

First-Order Theory Revision from Examples is the process of improving user-defined or automatically generated First-Order Logic (FOL) theories, given a set of examples. So far, the usefulness of Theory Revision systems has been limited by the cost of searching the huge search spaces they generate. This is a general difficulty when learning FOL theories but recent work showed that Stochastic Local Search (SLS) techniques may be effective, at least when learning FOL theories from scratch. Motivated by these results, we propose novel SLS based search strategies for First-Order Theory Revision from Examples. Experimental results show that introducing stochastic search significantly speeds up the runtime performance and improve accuracy.

One promising family of search strategies to alleviate runtime and storage requirements of ILP systems is that of stochastic local search methods, which have been successfully applied to hard propositional tasks such as satisfiability. Stochastic local search algorithms for propositional satisfiability benefit from the ability to quickly test whether a truth assignment satisfies a formula. Because of that many possible solutions can be tested and scored in a short time. In contrast, testing whether a clause covers an example in ILP takes much longer, so that far fewer possible solutions can be tested in the same time. Therefore in this paper we investigate stochastic local search in ILP using a relational propositionalized problem instead of directly use the first-order clauses space of solutions.

There has been significant recent progress in the integration of probabilistic reasoning with first order logic representations (SRL). So far, the learning algorithms developed for these models all learn from scratch, assuming an invariant background knowledge. As an alternative, theory revision techniques have been shown to perform well on a variety of machine learning problems. These techniques start from an approximate initial theory and apply modifications in places that performed badly in classification. In this work we describe the first revision system for SRL classification, PFORTE, which addresses two problems: all examples must be classified, and they must be classified well. PFORTE uses a two step-approach. The completeness component uses generalization operators to address failed proofs and the classification component addresses classification problems using generalization and specialization operators. Experimental results show significant benefits from using theory revision techniques compared to learning from scratch.

Recently, there has been great interest in integrating first-order logic based formalisms with mechanisms for probabilistic reasoning, thus defining probabilistic first-order theories (PFOT).Several algorithms for learning PFOTs have been proposed in the literature. They all learn the model from scratch. Consider a PFOT approximately correct, i.e., such that only a few points of its structure prevent it from reflecting the database correctly. It is much more efficient to identify these points and then propose modifications only to them than to use an algorithm that learns the theory from scratch. Therefore, in [3] we proposed a Bayesian Logic Programs Revision system (RBLP), which receives an initial BLP and through the examples discovers points that fail in covering some of them, similarly to FORTE [4] for the logical approach. These points are called logical revision points. RBLP then considers modifications only for those points choosing the best one through a scoring function. It is required that the implemented modification improves examples covering. It is expected that the returned BLP is consistent with the database.When learning or revising probabilistic first-order theories negative examples are incorporated into the set of positive examples, since the distributions of probabilities will reflect this difference in accordance with the domain of the predicates. At first, this would suggest only using generalization operators. The probabilistic learning algorithms also considers specialization operators where specialization is guided by the scoring function. The question arises of whether using specialization operators when revising a PFOT can improve classification and the result of the scoring function.In [2], besides experimentally comparing scoring functions, we extended RBLP, presenting PFORTE, arguing for the use of specialization operators even when there are no negative examples. We defined then probabilistic revision points, which are the places in theory that result in inaccurate classification of examples (the example was proved, but the value infered for the class is not the one given in the example). Modifications in these points are proposed by specialization operators and the best one is chosen through a scoring function. It is required that these modifications improve the score while not allowing any example to be unproved.Although the ideas presented here can be used for most kinds of PFOTs, we used BLP [1] to implement our system and experimentally compare the results.In the present work, we further study the benefits of considering specialization operators. We compare PFORTE, RBLP, and RBLP modified to allow specialization when a rule is being created by the add rule operator, using four datasets and considering conditional log likelihood as scoring function (since it obtained in [2] the best results). The resultant probabilistic accuracy for PFORTE was the best one considering p

Recently, there has been significant work in the integration of probabilistic reasoning with first order logic representations. Learning algorithms for these models have been developed and they all considered modifications in the entire structure. In a previous work we argued that when the theory is approximately correct the use of techniques from theory revision to just modify the structure in places that failed in classification can be a more adequate choice. To score these modifications and choose the best one the log likelihood was used. However, this function was shown not to be well-suited in the propositional Bayesian classification task and instead the conditional log likelihood should be used. In the present paper, we extend this revision system showing the necessity of using specialization operators even when there are no negative examples. Moreover, the results of a theory modified only in places that are responsible for the misclassification of some examples are compared with the one that was modified in the entire structure using three databases and considering four probabilistic score functions, including conditional log likelihood.