IDEAL 2018: 19TH INTERNATIONAL CONFERENCE ON INTELLIGENT DATA ENGINEERING AND AUTOMATED LEARNING
PROGRAM FOR FRIDAY, NOVEMBER 23RD
Days:
previous day
all days

View: session overviewtalk overview

10:00-11:30 Session 10C: Workshop on RiskTrack: Analyzing radicalization in Online Social Networks
Location: Meeting room
10:00
Introductory talk: Present and future perspectives on radicalization tracking
10:15
On Detecting Online Radicalization Using Natural Language Processing

ABSTRACT. In this paper, we propose a multi-facet based approach for identifying hate speech and extremism from Twitter dataset. Especially, building on previous research (Kostakos and Oussalah, 2018; Fong and Oussalah, 2017), we use various n-gram based features such as the presence of religious words, war-related terms and several hashtags that are commonly used in extremist tweets. Furthermore, other high-level linguistic cues like sentiment, personality change, emotion and emoticons as well as network related features are employed in order to grasp the rich and complexity of hate/extremism like text.Similarly to other related works we used a one-class classification framework involving advanced machine learning where SVM, K-NN, Random Forest were implemented. We conducted experiments on a real-world Twitter dataset and investigate the efficiency of the proposed approach.

10:30
RiskTrack: assessing the risk of Jihadi radicalization on Twitter using linguistic factors.
SPEAKER: Angel Panizo

ABSTRACT. RiskTrack is a project supported by the European Union, with the aim of helping security forces, intelligence services and prosecutors to assess the risk of Jihadi radicalization of an individual (or a group of people). To determine the risk of radicalization of an individual, it uses information extracted from her Twit-ter account. Specifically, the tool uses a combination of linguistic factors to estab-lish a risk value, in order to help the analyst with the decision making. This article aims to describe the linguistic features used on the first prototype of the Risk-Track tool. These factors, along with the way of calculating them and their contri-bution to the final risk value, will be presented in this paper. Also, some com-ments about the tool and the next updates will be suggested at the end of this pa-per.

10:45
Measuring Extremism: Validating an Alt-Right Twitter Accounts Dataset

ABSTRACT. Twitter is one of the most commonly used Online Social Networks in the world and it has consequently attracted considerable attention from different political groups attempting to gain influence. Among these groups is the alt-right; a mod-ern far-right extremist movement that gained notoriety in the 2016 US presiden-tial election and the infamous Charlottesville Unite the Right rally. This article de-tails the process used to create a database on Twitter of users associated with this movement, allowing for empirical research into this extremist group to be under-taken. In short, Twitter accounts belonging to leaders and groups associated with the Unite the Right rally in Charlottesville were used to create this database. After compiling users who followed these leading alt-right figures, an initial sample was gathered (n = 549). This sample was then validated by two researchers. This involved using a scoring method created for this process, in order to eliminate any accounts that were not supportive of the alt-right. Finally, a total amount of 422 accounts were found to belong to followers of this extremist movement, with a total amount of 123.295 tweets.

11:00
ONTOLOGY USES FOR RADICALISATION DETECTION on Social Networks

ABSTRACT. Social networks (SNs) are currently the main medium through which terrorist organisations reach out to vulnerable people with the objective of radicalizing and recruiting them to commit violent acts of terrorism. Fortunately, radicalization on social networks has warning signals and indicators that can be detected at the early stages of the radicalization process. In this paper, we explore the use of the semantic web and domain ontologies to automatically mine the radicalisation indicators from messages and posts on social networks.

11:15
Jihadi radicalized women in Spain: a profile

ABSTRACT. Although many previous articles have focused on assessing the profile of the jihadi terrorist or the jihadi radical, the role of the woman on the Jihad has been completely underestimated on the academic field. Women have get more and more prominent roles as part of the Jihadi plan, and therefore his role on the current conflict shall be understood as a part of the radical engine. The present article aims to present a general profile of the radicalized women in Spain, based on observations made with a sample of n = 21 subjects, all of them under vigi- lance or detained as part of a Jihadi cell. The information was gathered using open sources, information related to their criminal records and interviews to their families. All the outcomes obtained after the statistical assessment is presented on the results section, along with an explanation of the general profile of the ji- hadi radicalized woman in Spain.

10:30-11:30 Session 10A: Workshop on the Interplay between Human-Computer Interaction and Data Science (HCI-DS)
Location: Mixed room
10:30
Introductory talk: Clustering and visualization of high-dimensional and complex data
10:45
Exploratory study of the effects of cardiac murmurs on electrocardiographic-signal-based biometric systems

ABSTRACT. The process of distinguishing among human beings through the inspection of acquired data from physical or behavioral traits is known as biometric identification. Mostly, fingerprint- and iris-based biometric techniques are used. Nowadays, since such techniques are highly susceptible to be counterfeited, new biometric alternatives are explored mainly based on physiological signals and behavioral traits -which are useful not only for biometric identification purposes but may also play a role as a vital signal indicator. In this connection, the electrocardiographic (ECG) signals have shown to be a suitable approach. Nonetheless, their informative components (morphology, rhythm, polarization, and among others) can be affected by the presence of a cardiac pathology. Even more, some other cardiac diseases cannot directly be detected by the ECG signal inspection but still have an effect on their waveform, that is the case of cardiac murmurs. Therefore, for biometric purposes, such signals should be analyzed submitted to the effects of pathologies. This paper presents an exploratory study aimed at assessing the influence of the presence of a pathology when analyzing ECG signals for implementing a biometric system. For experiments, a database holding 20 healthy subjects and 20 pathological subjects (diagnosed with different types of cardiac murmurs) are considered. The proposed signal analysis consists of preprocessing, characterization (using wavelet features), feature selection and classification (five classifiers, as well as a mixture of them, are tested). As a result, through the performed comparison of the classification rates when testing pathological and normal ECG signals, the cardiac murmurs' undesired effect on the identification mechanism performance is clearly unveiled.

11:00
Model Selection in Committees of Evolved Convolutional Neural Networks using Genetic Algorithms

ABSTRACT. Neuroevolution is a technique that has been successfully applied for over three decades in order to optimize certain aspects of neural networks by applying evolutionary algorithms. However, only in the recent years, the increase of computational resources has enabled to apply such techniques to deep and convolutional neural networks, where the number of hyperparameters is significantly large.

In recent years, deep and convolutional neural networks are outperforming classical machine learning for many different tasks, including computer vision, natural language processing, signal processing, activity recognition, etc. In this context, neuroevolution can be useful since there are no analytic approaches for determining optimal network architectures or hyperparameters, therefore attaining better performance. Moreover, in some cases, committees (also called ensembles) are used, which combine two or more models in order to improve results even more. Neuroevolution can be of particular importance in this case, since evolutionary algorithms evolve a whole population of individuals, making it easy to build ensembles out of models chosen among those in the evolved population.

In this paper, we explore the application of genetic algorithms for carrying out model selection in a context of neuroevolution. Thus, the best models will be selected from a population of evolved individuals in order to maximize an overall objective function. This approach is tested using the well-known MNIST database as benchmark, and it obtains results which are highly competitive when compared with the state of the art.

11:15
Optimizing Meta-Heuristics for the Time-Dependent Traveling Salesman Problem Applied to Air Travels
SPEAKER: Diogo Duque

ABSTRACT. A travel agency has recently proposed the Traveling Salesman Challenge (TSC), a problem consisting of finding the best flights to visit a set of cities with the least cost. Our approach to this challenge consists on using a meta-optimized Ant Colony Optimization (ACO) strategy which, at the end of each iteration, generates a new ``ant'' by running Simulated Annealing or applying a mutation operator to the best ``ant'' of the iteration. Results are compared to variations of this algorithm, as well as to other meta-heuristic methods. They show that the developed approach is a better alternative than regular ACO for the time-dependent TSP class of problems, and that applying a K-Opt optimization in certain conditions may improve the results.

10:30-11:30 Session 10B: Big Data: Theory and applications
Chair:
10:30
Reusable Big Data System for Industrial Data Mining - A case study on Anomaly Detection in Chemical Plants

ABSTRACT. Industrial data mining projects in general and big data mining projects in particular suffer from long project execution. The resulting high costs render many otherwise interesting use cases economically unattractive. This contribution shows on the example of anomaly detection for process plant how the major obstacles - namely the inefficient development tools like Apache Hadoop and Spark and the lack of reuse of software artifacts across different projects can be overcome. This is achieved by selecting an application case that shares considerable commonalities across different projects and providing a supported project workflow implemented in a scalable and extensible big data architecture.

10:45
Spatial-Temporal K Nearest Neighbors Model on MapReduce for Traffic Flow Prediction

ABSTRACT. Researches in the area of short-term traffic flow forecasting are important for traffic flow management in intelligent transport systems (ITS). The results of short-term traffic flow forecasting in ITS may be used for different purposes: planning the optimal route, reducing the travel time, saving car’s fuel, alleviating roads congestion and so on. In this paper, a distributed model for short-term traffic flow prediction based on the k nearest neighbors method is presented. This model takes into account spatial and temporal traffic flow distribution. We define a feature vector for a targeted road segment using traffic flow on segments in a compact area at different time intervals. To reduce the dimensionality of the feature vector, we use principal component analysis procedure. The proposed model is based on MapReduce technology and implemented using an Apache Spark framework. An experimental study data is obtained from the transportation network of Samara, Russia.

11:00
Low Level Big Data Processing

ABSTRACT. The machine learning algorithms, prior to their application, require that the information be stored in memory. Reducing the amount of memory used for data representation clearly reduces the number of operations required to process it. Many of the current libraries represent the information in the traditional way, which forces you to iterate the whole set of data to obtain the desired result. In this paper we propose a technique to process categorical information previously encoded using the bit-level schema, the method proposes a block processing which reduces the number of iterations on the original data and, at the same time, maintains a processing performance similar to the processing of the original data. The method requires the information to be stored in memory, which allows you to optimize the volume of memory consumed for representation as well as the operations required to process it. The results of the experiments carried out show a slightly lower time processing than the obtained with traditional implementations, which allows us to obtain a good performance.

11:30-12:00Coffee Break
12:00-13:00 Session 11: Plenary talk
12:00
Statistical and Algorithmic Learning

ABSTRACT. Let's assume that the training set of our data follows the assumption of randomness - the examples are independent and identically distributed (i.i.d.). That is our only assumption and we want to check out if a new example follows this i.i.d. assumption. To estimate if the new example is a "typical" or "strange" in some sense (or how well it fits or conforms to the training set distribution), we need to define a strangeness (non-conformity) measure (NCM). To make a prediction or to estimate a confidence of the prediction we can convert our NCMs into statistical p-values using the Martin-Löf test for randomness. The highest p-value gives us a prediction and the (1 - 2nd highest p-value) gives us a confidence in the prediction. The learning machines based on these ideas is called conformal predictors (CP) and has a number of advantages over the conventional machine learning algorithms. In particular, CPs can be used in connection with any machine learning algorithms (support vector machine, neural networks/deep learning, kernel methods, random forest, logistic regression, etc.): we extract our measure of non-conformity from the underlying machine learning algorithm and use it in the calculations of p-values. It also connects a confidence with accuracy of the prediction. In other words, unlike in the classical machine learning algorithms, we can control a number of errors by setting up a required confidence level. In the online mode, we can also prove that the probability of error at every step is epsilon and errors are made independently at different trials. In combination with the law of large numbers, this implies “validity” property: the long-run percentage of the erroneous predictions will be close, with high probability, to the chosen significance level.

13:00-14:00Lunch Break
14:10-16:10 Session 12A: Social and Human Applications of Artificial Intelligence
14:10
Finding the Importance of Facial Features in Social Trait Perception

ABSTRACT. We are constantly making very fast attributions from faces, such as whether a person is trustworthy or threatening, that influence our behavior towards people. In this work, we present a method to automatically tell the importance of facial features on social trait perception. We employ an unsupervised clustering method to group the facial features by similarity and then create a model which explains the contribution of each facial feature to each social trait by means of a Genetic Algorithm. Our model deals with the difficulties associated to quantifying social perception using judgments from human observers (low inter-observer and intra-observer agreement) and obtains significant correlations greater than 0.7 for all social impressions, which justifies the method followed. Finally, weights of eyebrows, eyes, nose, mouth, jawline and facial feature distances are shown and discussed. This work poses a step forward in social trait perception understanding, as to the date, there is no other work quantifying the effects of facial features on social trait perception.

14:25
Extending Independent Component Analysis on Online Social Media for Event Detection
SPEAKER: Jason Jung

ABSTRACT. Social networking service is a good sensor to reflect social events because people tend to publish everything online. With the support of smart devices, events on social networking service are even real-time because users can post messages at any time in any place without any barrier. One of the greatest features of some social networking services (e.g., Twitter, Instagram) is that they provide hashtags. As an indicator, Twitter hashtags assist users to generate events in an advantageous way. However, a wrong behavior of using hashtags is creating unavailing data and reducing this kind of noise is considered as an essential task. In this paper, we propose a new approach for filtering noise from social event signals by using independent component analysis technique. With a case study about two practical events, we prove that our approach is feasible. Besides, this idea is also to acquire high adaptability and extensibility property because it uses only temporal signals of social events as inputs without requiring any more information. Moreover, outputs of our approach can also be used as pre-processing step for other signal-based event discovering algorithms.

14:40
Chatbot Theory: A Naı̈ve and Elementary Theory for Dialogue Management

ABSTRACT. Due to the increasing interested and use of chatbot, its properties and operation possibilities shall be proper realized matching both safety and security issues as well as present the several uses and compositions that this technology supports. This paper focus is on dialogue management since it is considered the core of a chatbot. The dialogue manager is responsible to, more than to transform an input sentence into an output one, hold the illusion of a human conversation. In this sense, it is presented an inceptive theoretical framework through a formal way for chatbots that can be used as a reference to explore, compose, build and discuss chatbots. The discussion is performed mostly on ELIZA since, due to its historical records, it can be considered an important reference chatbot, nevertheless, the proposed theory is compatible with the most recent technologies such those using machine and deep learning. The paper then presents some sketchy instances in order to explore the support provided by the theory.

14:55
Understanding Learner's Drop-out in MOOCs

ABSTRACT. This paper focuses on anticipating the drop-out among MOOC learners and helping in the identification of the reasons behind this drop-out. The main reasons are those related to course design and learners behavior, according to the requirements of the MOOC provider OpenClassrooms. Two critical business needs are identified in this context. First, the accurate detection of at-risk droppers, which allows sending automated motivational feedback to prevent learners drop-out. Second, the investigation of possible drop-out reasons, which allows making the necessary personalized interventions. To meet these needs, we present a supervised machine learning based drop-out prediction system that uses \emph{Predictive algorithms} (Random Forest and Gradient Boosting) for automated intervention solutions, and \emph{Explicative algorithms} (Logistic Regression, and Decision Tree) for personalized intervention solutions. The performed experimentations cover three main axes; (1) Implementing an enhanced reliable dropout-prediction system that detects at-risk droppers at different specified instants throughout the course. (2) Introducing and testing the effect of advanced features related to the trajectories of learners' engagement with the course (backward jumps, frequent jumps, inactivity time evolution). (3) Offering a preliminary insight on how to use readable classifiers to help determine possible reasons for drop-out. The findings of the mentioned experimental axes prove the viability of reaching the expected intervention strategies.

15:10
Unsupervised Domain Adaptation for Human Activity Recognition

ABSTRACT. Human Activity Recognition has been largely investigated as a machine learning classification task. Human Activity Recognition systems usually assume that each activity is carried out in the same way by different humans. However, a specific person can perform physical activities differently from others, and even sensors are prone to misfunction. However, to model the pattern of activities carried out by each user, a significant amount of data is needed. Thus, data must be collected through a long period of time. To reduce this problem, this paper investigates the use of Transfer Learning, specifically Unsupervised Domain Adaptation, within human activity recognition systems. The yielded experiment results reveal a useful transfer of knowledge and more importantly the convenience of transfer learning within human activity recognition. Apart from the delineated experiments, our work also contributes to the field of transfer learning in general through an exhaustive survey on transfer learning for human activity recognition based on wearables.

15:25
Communication Skills Personal Trainer based on Viola-Jones object detection algorithm

ABSTRACT. The paper presents a personal training system that allows the user to use non-verbal communication techniques that help him improve his way of speaking in public. The tool is a software solution that implements algorithms to identify ele-ments of non-verbal communication (NVC), such as the positions of the head, hands and trunk of the users making use of cascading classifiers. It presents the developed technology and the different characteristics of the application that is an important part of the treatment in artificial vision techniques in the detection of NVC and its training

15:40
Predicting Online Review Scores Across Reviewer Categories

ABSTRACT. In this paper, we propose and test an approach based on regression models, to predict the review score of an item, across different reviewer categories.

The analysis is based on a public dataset with more than 2,5 million hotel reviews, belonging to five specific reviewers' categories. We first compute the relation between the average scores associated with the different categories and generate the corresponding regression model. Then, the extracted model is used for prediction: given the average score of a hotel according to a reviewer category, it predicts the average score associated with another category.

15:55
Generalized low-computational cost Laplacian eigenmaps

ABSTRACT. Dimensionality reduction (DR) is a methodology used in many fields linked to data processing, and may represent a preprocessing stage or be an essential element for the representation and classification of data. The main objective of DR is to obtain a new representation of the original data in a space of smaller dimension, such that more refined information is produced, as well as the time of the subsequent processing is decreased and/or visual representations more intelligible for human beings are generated. The spectral DR methods involve the calculation of an eigenvalue and eigenvector decomposition, which is usually high-computational-cost demanding, and, therefore, the task of obtaining a more dynamic and interactive user-machine integration is difficult. Therefore, for the design of an interactive IV system based on DR spectral methods, it is necessary to propose a strategy to reduce the computational cost required in the calculation of eigenvectors and eigenvalues. For this purpose, it is proposed to use locally linear submatrices and spectral embedding. This allows integrating natural intelligence with computational intelligence for the representation of data interactively, dynamically and at low computational cost. Additionally, an interactive model is proposed that allows the user to dynamically visualize the data through a weighted mixture.

14:10-16:10 Session 12B: Deep learning & Neural Networks
Location: Mixed room
14:10
Deep Learning applied to Wind Energy Forecasting

ABSTRACT. One of the ways to reduce the effects of Climate Change is to rely in energy generation based on renewable sources. Their intermittent nature makes necessary to obtain a mid-long term accurate forecasting. Wind Energy prediction is based on the ability to forecast wind speed. This has been a problem approached using different methods based on the statistical properties of the wind time series with meteorological prediction inputs. In this paper we focus in prediction made without integration of weather forecasts. Wind Time series are non-linear and non-stationary, making their forecasting very challenging. Deep neural networks have shown their success recently for problems involving sequences with non-linear behavior. In this work, we perform experiments comparing the capability of different neural network architectures for multi-step forecasting in a 12 hour ahead prediction. For the Time Series input we used the US National Renewable Energy Laboratory’s WIND Dataset, (the largest available wind and energy dataset with over 120.000 physical sites), this dataset is evenly spread across all the North America geography which has allowed us to obtain conclusions on the relationship between physical site complexity and forecast accuracy.

14:25
An approximation to Deep Learning touristic-related time series forecasting

ABSTRACT. Tourism is one of the biggest economic activities around the world. This means that an adequate planning of existing resources becomes crucial. Precise demand-related forecasting greatly improves this planning. \emph{Deep Learning} models are showing an greatly improvement on time-series forecasting, particularly the \emph{LSTM}, which is designed for this kind of tasks. This article introduces the touristic time-series forecasting using LSTM, and compares its accuracy against well known models RandomForest and ARIMA.

Our results shows that new LSTM models achieve the best accuracy.

14:40
Towards Complex Features: Competitive Receptive Fields in Unsupervised Deep Networks
SPEAKER: Hujun Yin

ABSTRACT. In this paper we propose a simple unsupervised approach to learn higher order features. This model is based on the recent success of lightweight approaches such as SOMNet and PCANet to the challenging task of image classification. Contrary to the more complex deep learning models such as convolutional neural networks (CNNs), these methods use naive algorithms to model the input distribution. Our endeavour focuses on the self-organizing map (SOM) based method and extends it by incorporating a competitive connection layer between filter learning stages. This simple addition encourages the second filter learning stage to learn complex combinations of first layer filters and simultaneously decreases channel depth. This approach to learning complex representations offers a competitive alternative to common deep learning models whilst maintaining an efficient framework. We test our proposed approach on the popular MNIST and challenging CIFAR-10 datasets.

14:55
Framework for the Training of Deep Neural Networks in TensorFlow using Metaheuristics

ABSTRACT. Artificial neural networks (ANN) again are playing a leading role in the area of machine learning, especially in classification and regression processes, due to the emergence of deep learning (ANNs with more than four hidden layers), allowing them to encode more and more complex features. The increase in the number of hidden layers in ANNs has posed important challenges in their training. Variations (eg RMSProp) of classical algorithms such as backpropagation with its stochastic gradient descent are the state of the art for training deep ANNs. However, other research has shown that the advantages of metaheuristics need more detailed study in this area. We summarize the design and use of a framework to optimize learning of deep neural networks in TensorFlow using metaheuristics, a framework implemented in Python that allows training of the networks in CPU or GPU depending on the TensorFlow configuration and allows easy integration of diverse classification and regression problems solved with different neural networks architectures (conventional, convolutional and recurrent) and new metaheuristics. The framework initially includes Particle Swarm Optimization, Global-best Harmony Search, and Differential Evolution. It further enables the conversion of metaheuristics into memetic algorithms including exploitation processes using the algorithms available in TensorFlow: RMSProp, Adam, Adadelta, Momentum and Adagrad.

15:10
Improving SeNA-CNN by Automating Task Recognition
SPEAKER: Abel Zacarias

ABSTRACT. Catastrophic forgetting arises when a neural network is not capable of preserving the past learned task when learning a new task. There are already some methods proposed to mitigate this problem in artificial neural networks. In this paper we propose to improve upon our previous state-of-the-art method, SeNA-CNN, such as to enable the automatic recognition in test time of the task to be solved and we experimentally show that it has excellent results. The experiments show the learning of up to 4 different tasks with a single network, without forgetting how to solve previous learned tasks.

15:25
Learning Optimal Q-function using Deep Boltzmann Machine for Reliable Trading of Cryptocurrency
SPEAKER: Seok-Jun Bu

ABSTRACT. The explosive price volatility from the end of 2017 to January 2018 shows that bitcoin is a high risk asset. The deep reinforcement algorithm is straightforward idea for directly outputs the market management actions to achieve higher profit instead of higher price-prediction accuracy. However, existing deep reinforcement learning algorithms including Q-learning are also limited to problems caused by enormous searching space. We propose a combination of double Q-network and unsupervised pre-training using Deep Boltzmann Machine (DBM) to generate and enhance the optimal Q-function in cryptocurrency trading. We obtained the profit of 2,686% in simulation, whereas the best conventional model had that of 2,087% for the same period of test. In addition, our model records 24% of profit while market price significantly drops by -64%.

15:40
Detecting Intrusive Malware with a Hybrid Generative Deep Learning Model
SPEAKER: Jin-Young Kim

ABSTRACT. A small amount of unknown malware can be analyzed manually, but it is generated with extremely more and more so that automatic detection of them is needed. Malware is usually generated with different features from those of existing ones (e.g., code exchange, null value insertion, or reorganization of subroutines) to avoid detection of antivirus systems. To detect malware with obfuscation, this paper proposes a method called latent semantic controlling generative adversarial networks (LSC-GAN) that learns to generate malware data with i-feature from a specific Gaussian distribution which represents i-feature and distinguish it from the real. Variational autoencoder (VAE) projects data to latent space for feature extraction and is transferred to generator (G) of LSC-GAN to train it stably. G generates data from Gaussian distribution, so it produces similar data but not identical to the actual data: it includes modified features compared with the real. The detector is inherited with transfer learning in a encoder that learns various malware features using real and modified data generated by the LSC-GAN based on a LSC-VAE. We show that LSC-GAN achieves detection accuracy of 96.97% on average that is higher than those of other conventional models. We demonstrate statistical significance of the performance of the proposed model using t-test. The result of detection is analyzed with confusion matrix and F1-score.

15:55
Predicting the Household Power Consumption using CNN- LSTM Hybrid Networks
SPEAKER: Tae-Young Kim

ABSTRACT. Power consumption prediction is an essential element of the electricity supply company's operation and development plan. In terms of power supply and de-mand planning, reserve power must always be prepared to provide stable power to consumers. However, since electricity cannot be stored, it is necessary to fore-cast electricity demand and produce electricity. In this paper, we propose a CNN-LSTM hybrid neural network that can extract spatial and temporal feature to ef-fectively predict the house power consumption. Experiments have shown that CNN-LSTM hybrid network, which combines convolutional neural network (CNN), long short-term memory (LSTM) and deep neural network (DNN), can extract complex features of electric power consumption. The CNN layer is used to reduce the noise of spatial information, the LSTM layer is suitable for model-ing temporal information, and the DNN layer maps normal and abnormal data in-to different spaces. Our CNN-LSTM hybrid method achieves almost perfect pre-diction performance for electric power consumption that was previously difficult to predict. Finally, the CNN LSTM hybrid method achieves higher RMSE per-formance than other state-of-the-art predicting methods in individual household electric power consumption datasets provided by UCI.

14:10-15:10 Session 12C-I: Special Session on New Models of Bio-inspired computation for Massive Complex Environments (NMBC)
Location: Meeting room
14:10
A metaheuristic approach for the α-separator problem

ABSTRACT. Most of the critical infrastructures can be easily modeled as a network of nodes interconnected among them. If one or more nodes of the network fail, the connectivity of the network can be compromised, to the point of completely disconnecting the network. Additionally, disconnecting the network can result in cascade failures, because the remaining nodes may be overloaded because of heavy traffic in the network. One of the main objectives of an attacker is to isolate the nodes whose removal disconnect the network in minimum size subnetworks. On the contrary, a defender must identify those weak points in order to maintain the network integrity. This work is focused on solving the α separator problem, whose main objective is to find a minimum set of nodes that disconnect a network in isolated subnetworks of size smaller than a given value. The problem is tackled from a metaheuristic point of view, analyzing the solutions given by a Greedy Randomized Adaptive Search Procedure over different network topologies. The results obtained are compared with the best algorithm found in the literature.

14:25
Community Detection in Weighted Directed Networks using Nature-inspired Heuristics
SPEAKER: Eneko Osaba

ABSTRACT. Finding groups from a set of interconnected nodes is a recurrent paradigm in a variety of practical problems that can be modeled as a graph, as those emerging from Social Networks. However, finding an optimal partition of a graph is a computationally complex task, calling for the development of approximative heuristics. In this regard, the work presented in this paper tackles the optimal partitioning of graph instances whose connections among nodes are directed and weighted, a scenario significantly less addressed in the literature than their unweighted, undirected counterparts. To efficiently solve this problem, we design several heuristic solvers inspired by different processes and phenomena observed in Nature (namely, Water Cycle Algorithm, Firefly Algorithm, an Evolutionary Simulated Annealing and a Population based Variable Neighborhood Search), all resorting to a reformulated expression for the well-known modularity function to account for the direction and weight of edges within the graph. Extensive simulations are run over a set of synthetically generated graph instances, aimed at elucidating the comparative performance of the aforementioned solvers under different graph sizes and levels of intra- and inter-connectivity among node groups. We statistically verify that the approach relying on the Water Cycle Algorithm outperforms the rest of heuristic methods in terms of Normalized Mutual Information with respect to the true partition of the graph.

14:40
An Artificial Bee Colony algorithm for optimizing the design of sensor networks

ABSTRACT. The sensor network design problem (SNDP) consists of the selection of the type, number and location of the sensors to measure a set of variables, optimizing a specified criteria, and simultaneously satisfying the information requirements. This problem is multimodal and involves several binary variables, therefore it is a complex combinatorial optimization problem. This paper presents a new Artificial Bee Colony (ABC) algorithm designed to solve high scale designs of sensor networks. For this purpose, the proposed ABC algorithm has been designed to optimize binary structured problems that also handling constraints to fulfill information requirements. The classical version of ABC algorithm was proposed for solving unconstrained and continuous optimization problems. Several extensions have been proposed that allow the classical ABC algorithm to work on constrained or on binary optimization problems. Therefore the proposed approach is a new version of the ABC algorithm that combines the binary and constrained optimization extensions to solve the SNDP. Finally the new algorithm is tested using different systems of incremental size to evaluate its quality, robustness, and scalability.

14:55
Design of Japanesse Tree Frog algorithm for Community Finding Problems

ABSTRACT. Community Finding Problems (CFP) have become very popular in the last years, due to the high number of users that connect everyday to Social Networks (SNs). The goal of these problems is to group the users that compose the SN in several clusters, or circles, in such a way similar users belong to the same cluster, whereas different users are assigned to different clusters. Due to the high complexity of this problem, it is common that researchers use heuristic algorithms to perform this task in a reasonable computational time. This paper is focused on the applicability of a novel bio-inspired algorithm to solve CFPs. The selected algorithm is based on the real behaviour of the Japanesse Tree Frog, that has been sucessfully used to color maps and extract the Maximal Independet Set of a graph.

15:10-16:10 Session 12C-II: Recommender Systems
Location: Meeting room
15:10
Taking e-Assessment Quizzes - A Case Study with an SVD Based Recommender System

ABSTRACT. Recommending learning assets in e-Learning systems represents a key aspect. Among many available assets there are quizzes that validate and also evaluate learner's knowledge level. This paper presents a recommender system based on SVD algorithm that is able to properly recommend quizzes such that learner's knowledge level is evaluated and displayed in real time by means of a custom designed concept map for graph algorithms within the Data Structures course. A preliminary case study presents a comparative analysis between a group a learners that received random quizzes and a group of learners that received recommended questions. The visual analytics and interpretation of two representative cases show a clear advantage of the students received recommended questions over the other ones.

15:25
Exploring coclustering for serendipity improvement in content-based recommendation

ABSTRACT. Content-based recommender systems are now widely used for item recommendations in several application domains such as entertainment, e-commerce and news. However, one of its major drawbacks is the lack of serendipity in recommendations. A recommendation is considered serendipitous when is both relevant and unexpected. There is a common understanding in literature that the search for serendipitous recommendations should be guided by partial similarities between items. From that intuition, coclustering can be exploited in order to ensure the compromise between accuracy and unexpectedness leading to serendipitous recommendations to the users, since it is a technique capable of finding partial similarity relations between items. In this paper, we propose the use of coclustering for serendipity improvement in content-based recommender systems. Experiments carried out over the MovieLens 2K dataset show that the proposed approach overcome a traditional content-based recommender in terms of serendipity.

15:40
An Improved Fuzzy Logic Based Recommender System by Integrating Social Tags and Social Networks’ Information

ABSTRACT. With the rapid development of computers, internet, social media and networks, and other digital multimedia technologies, it is needed to use a mechanism that can predict the needs and desires of users and recommend the bests for them. Introducing the social networks’ information into the traditional collaborative filtering (CF) algorithm, the essay studies the changes of user preference in social networks. Recently a lot of research efforts have been spent on building recommender systems by utilizing the abundant online social network’s data. This paper proposes an improved collaborative filtering algorithm based on fuzzy logic and Social Network Information. The proposed method enhances the accuracy of recommendations by combining the social tags, fuzzy logic and social networks’ information such as friendship and groups’ membership. Through the experiment, the improved algorithm has higher accuracy than the traditional filtering algorithms in the top-N recommendation list. It proves that the social networks’ information of users can affect the user's preference.

15:55
Novel and Classic Metaheuristics for Tunning a Recommender System for Predicting Student Performance in Online Campus

ABSTRACT. The prediction of the students' performance allows to improve the learning process using the online campus tools. In this context, recommender systems are useful for prediction purposes. This collaborative filtering tool, predicts the unknown performances analyzing the database that contains the performance of the students for particular tasks, considering matrix factorization and stochastic gradient descent. If we consider a fixed number of latent factors, the prediction error is mainly influenced by two parameters: learning rate and regularization factor. The best settings for these parameters is an optimization problem that can be tackled by soft computing techniques. In this work, we analyze three solving methods to select the optimal values of both parameters: a simple direct search, a classic evolutionary algorithm, and a novel metaheuristic. The results show the advantages of using metaheuristics instead of direct search in accuracy and computing effort terms.