Eduardo Bezerra

A Forecasting Method for Fertilizers Consumption in Brazil

Tracking information about fertilizers consumption in the world is very important since they are used to produce agriculture commodities. Brazil consumes a large amount of fertilizers due to its large-scale agriculture fields. Most of these fertilizers are currently imported. The analysis of consumption of major fertilizers, such as Nitrogen-Phosphorus-Potassium (NPK), Sulfur, Phosphate Rock, Potash, and Nitrogen become critical for long-term government decisions. In this paper we present a method for fertilizers consumption forecasting based on both Autoregressive Integrated Moving Average (ARIMA) and logistic function models. Our method was used to forecast fertilizers consumption in Brazil for the next 20 years considering different economic growth for the entire country.

Semi-Supervised Clustering of XML Documents: Getting the Most from Structural Information

As document providers can express more contextualized and complex information, semi-structured documents are becoming a major source of information in many areas, e.g., in digital libraries, e-commerce or Web applications. A particular characteristic of such document collections is the existence of some structure or metadata along with the data. In this scenario, clustering methods that can take advantage of such structural information to better organize such collections are highly relevant. Semi-structured documents pose new challenges to document clustering methods, however, since it is not clear how this structural information can be used to improve the quality of the generated clustering models. On the other hand, recently there has a growing interest in the semi-supervised clustering task, in which a little amount of prior knowledge is provided to guide the algorithm to a better clustering model. A particular type of semi-supervision is in the form of user-provided constraints defined over pairs of objects, where each pair informs if its objects must be in the same or in different clusters. In this paper, we consider the problem of constrained clustering in documents that present some form of structural information. We consider the existence of a particular form of information to be clustered: textual documents that present a logical structure represented in XML format. We define and extend methods to improve the quality of clustering results by using such structural information to guide the execution of the constrained clustering algorithm. Experimental results on the OHSUMED document collection show the effectiveness of our approach.

A framework for benchmarking machine learning methods using linear models for univariate time series prediction

Time series prediction has been attracting interest of researchers due to its increasing importance in decision-making activities in many fields of knowledge. The demand for better accuracy in time series prediction furthered the arising of many machine learning time series prediction methods (MLM). Choosing a suitable method for a particular dataset is a challenge and demands established benchmark methods (BM) for performance assessment. Suppose a particular BM is selected, and an experimental comparison is made with a particular MLM. If the latter does not provide better prediction results for the same dataset, this indicates that some improvements are needed for the MLM. Regarding this matter, adopting a well-established, easy to interpret, and tuned BM is desirable. This paper presents a framework for systematic benchmarking some MLM against well-known Linear Methods (LM), namely Polynomial Regression and models in the ARIMA family, used as BM for univariate time series prediction. We implemented such a framework within the R-Package named TSPred. This implementation was evaluated using a wide number of datasets from past prediction competitions. The results show that fittest LM provided by TSPred are adequate BM for univariate time series predictions.

Exploring multiple clusterings in attributed graphs

Many graph clustering algorithms aim at generating a single partitioning (clustering) of the data. However, there can be many ways a dataset can be clustered. From a exploratory analisys perspective, given a dataset, the availability of many different and non-redundant clusterings is important for data understanding. Each one of these clusterings could provide a different insight about the data. In this paper, we propose M-CRAG, a novel algorithm that generates multiple non-redundant clusterings from an attributed graph. We focus on attributed graphs, in which each vertex is associated to a n-tuple of attributes (e.g., in a social network, users have interests, gender, age, etc.). M-CRAG adds Artificial edges between similar vertices of the attributed graph, which results in an augmented attributed graph. This new graph is then given as input to our clustering algorithm (CRAG). Experimental results indicate that M-CRAG is effective in providing multiple clusterings from an attributed graph.

A Crowdsourcing Tool for Data Augmentation in Visual Question Answering Tasks

Visual Question Answering (VQA) is a task that connects the fields of Computer Vision and Natural Language Processing. Taking as input an image I and a natural language question Q about I, a VQA model must be able to produce a coherent answer R (also in natural language) to Q. A particular type of visual question is one in which the question is binary (i.e., a question whose answer belongs to the set {yes, no}). Currently, deep neural networks correspond to the state of the art technique for training of VQA models. Despite its success, the application of neural networks to the VQA task requires a very large amount of data in order to produce models with adequate precision. Datasets currently used for the training of VQA models are the result of laborious manual labeling processes (i.e., made by humans). This context makes relevant the study of approaches to augment these datasets in order to train more accurate prediction models. This paper describes a crowdsourcing tool which can be used in a collaborative manner to augment an existing VQA dataset for binary questions. Our tool actively integrates candidate items from an external data source in order to optimize the selection of queries to be presented to curators.

A Mixed Graph Framework to evaluate the complementarity of communication Tools.

18 Due to the constant innovations in communications tools, several organizations are constantly evaluating the adoption of new communication tools (NCT) with respect to current ones. Especially, many organizations are interested in checking if NCT is really bringing benefits in their production process. We can state an important problem that tackles this interest as for how to identify when NCT is providing a significantly different complementary communication flow with respect to the current communication tools (CCT). This paper presents the Mixed Graph Framework (MGF) to address the problem of measuring the complementarity of a NCT in the scenario where some CCT is already established. We evaluated MGF using synthetic data that represents an enterprise social network (ESN) in the context of well-established e-mail communication tool. Our experiments observed that the MGF was able to identify whether a NCT produces significant changes in the overall communications according to some centrality measures. 19 20 21 22 23 24 25 26 27 28

Exploring machine learning methods for the Star/Galaxy Separation Problem.

For recent or planned deep astronomical surveys, it is important to tell stars and galaxies apart, a task known as Star/Galaxy Separation Problem (SGSP). At faint magnitudes, the separation between pointy and extended sources is fuzzy, which makes SGSP a hard task. This problem is even harder for large surveys like Dark Energy Survey (DES) and, in a near future, the Large Synoptic Survey Telescope (LSST) due to their large data volume. Hence, the search for classification methods that are both accurate and efficient is highly relevant. In this work, we present a comparative analysis of several machine learning methods targeted at solving the SGSP at faint magnitudes. In order to train the classification models, the COSMOS survey was used. We use machine learning methods as distinct as artificial neural networks, k nearest-neighbor, Support Vector Machines, Random Forests and Naive Bayes. The exploratory process was modeled as data centric workflow. The workflow was implemented on top of Hadoop framework and was used to find the best parameter values for each classification method we considered, of which neural networks and random forest present superior performance.

Evaluating temporal aggregation for predicting the sea surface temperature of the Atlantic Ocean

Abstract Extreme environmental events such as droughts affect millions of people all around the world. Although it is not possible to prevent this type of event, its prediction under different time horizons enables the mitigation of eventual damages caused by its occurrence. An important variable for identifying occurrences of droughts is the sea surface temperature (SST). In the tropical Atlantic Ocean, SST data are collected and provided by the Prediction and Research Moored Array in the Tropical Atlantic (PIRATA) project, which is an observation network composed of sensor buoys arranged in this region. Sensors of this type, and more generally Internet of Things (IoT) sensors, commonly lead to data losses that influence the quality of datasets collected for adjusting prediction models. In this paper, we explore the influence of temporal aggregation in predicting step-ahead SST considering different prediction horizons and different sizes for training datasets. We have conducted several experiments using data collected by PIRATA project. Our results point out scenarios for training datasets and prediction horizons indicating whether or not temporal aggregated SST time series may be beneficial for prediction.

Amê: an environment to learn and analyze adversarial search algorithms using stochastic card games

Computer Science students are usually enthusiastic about learning Artificial Intelligence (AI) due to the possibility of developing computer games that incorporate AI behaviors. Under this scenario, Search Algorithms (SA) are a fundamental subject of AI for a broad variety of games. Implementing deterministic games, varying from tic-tac-toe to chess games, are commonly approaches used to teach AI. Considering the perspective of game playing, however, stochastic games are usually more fun to play, and are not much explored during AI learning process. Other approaches in AI learning include developing searching algorithms to compete against each other. These approaches are relevant and engaging, but they lack an environment that features both algorithm design and benchmarking capabilities. To address this issue, we present Amê -- an environment to support the learning process and analysis of adversarial search algorithms using a stochastic card game. We have conducted a pilot experiment with Computer Science students that developed different adversarial search algorithms for Hanafuda (a traditional Japanese card game).