Salma Jamoussi

Weighted ensemble learning of Bayesian network for gene regulatory networks

Abstract Gene Regulatory Network (GRN) is known as the most adequate representation of genes׳ interactions based on microarray datasets. One of the most performing modeling tools that enable the inference of these networks is a Bayesian network (BN). When preceded by an efficient pre-processing step, BN learning can unveil possible relationships between key disease genes and allows biologists to analyze these interactions and to exploit them. However, the layout of microarray data is different from classic data. This particularity engenders challenges to BN learning in terms of dimensionality and data over-fitting. In this paper, we propose a fuzzy ensemble clustering method that allows outputting small and highly inter-correlated partitions of genes so that we can overcome dimensionality problem. We present a weighted committee based structure algorithm for learning BNs of each partition without over-fitting training dataset. Moreover, we offer an approach for assembling the sub-BNs through genes in common. We also statistically verify and biologically validate our approach.

Dynamic construction of dictionaries for sentiment classification

The sentiment classification is one of the new challenges emerged with the advence of social networks. Our purpose is to determine the sentimental orientation of a Facebook comment (positive or negative) by using the linguistic approach. In most of the sentiment analysis applications using this approach, the sentiment lexicon plays a key role. Thus, it is very important to create a lexicon covering several sentiment words. For this reason, we address in this paper the problem how to group and list words present in the corpus into two dictionaries. We proposed a new automatic technique to create the positive and negative dictionaries that exploits the emotions symbols (emoticons, acronyms and exclamation words) present in comments. More importantly, our idea allows to enlarge these dictionaries with an enrichment step. Finally, by using these prepared dictionaries, we predict the positive and negative polarities of the comment. We evaluate our approach by comparison to human classification. Our results are also effective and consistent.

Principal component analysis neural network for textual document categorization and dimension reduction

This manuscript presents the study and application of the method of principal component analysis (PCA) in the field of text mining. We began by studying the theoretical basis behind this method and we have focused on two of its variants namely the neural PCA and kernel PCA. We used neural PCA for automatic categorization of text documents through an extraction of semantic concepts. The second contribution of our work is the use of PCA (neuronal and kernel) for the dimension reduction of textual documents through the automatic classification.

Dynamic Construction of Dictionaries for Sentiment Classification

The sentiment classification is one of the new challenges emerged with the advence of social networks. Our purpose is to determine the sentimental orientation of a Facebook comment (positive or negative) by using the linguistic approach. In most of the sentiment analysis applications using this approach, the sentiment lexicon plays a key role. Thus, it is very important to create a lexicon covering several sentiment words. For this reason, we address in this paper the problem how to group and list words present in the corpus into two dictionaries. We proposed a new automatic technique to create the positive and negative dictionaries that exploits the emotions symbols (emoticons, acronyms and exclamation words) present in comments. More importantly, our idea allows to enlarge these dictionaries with an enrichment step. Finally, by using these prepared dictionaries, we predict the positive and negative polarities of the comment. We evaluate our approach by comparison to human classification. Our results are also effective and consistent.

RelANE: Discovering Relations between Arabic Named Entities

In this paper, we describe the first tool that detects the semantic relation between Arabic named entities, henceforth RelANE. We use various supervised learning techniques to predict the word or the sequence of terms that can highlight one or more semantic relationship between two Arabic named entities.

Grouping Like-Minded Users Based on Text and Sentiment Analysis

With the growth of social media usage, the study of online communities and groups has become an appealing research domain. In this context, grouping like-minded users is one of the emerging problems. Indeed, it gives a good idea about group formation and evolution, explains various social phenomena and leads to many applications, such as link prediction and product suggestion. In this dissertation, we propose a novel unsupervised method for grouping like-minded users within social networks. Such a method detects groups of users sharing the same interest centers and having similar opinions. In fact, the proposed method is based on extracting the interest centers and retrieving the polarities from the user’s textual posts.

A Probabilistic Ontology for the Prediction of Author’s Interests

The Bayesian network, a probabilistic model of knowledge representation, has the ability to represent and reason with uncertainty. It measures the dependencies between a set of variables and infer new knowledge. In this paper, we try to propose a method for building a probabilistic ontology, which models a list of publications (dblp base). We have used for this aim a Bayesian Network to measure dependencies between different instances of ontology and to infer new interests of authors from obtained Probabilistic Ontology.

Enhancing Machine Learning Results for Semantic Relation Extraction

This paper describes a large scale method to extract semantic relations between named entities. It is characterized by a large number of relations and can be applied to various domains and languages. Our approach is based on rule mining from an Arabic corpus using lexical, semantic and numerical features.

A New Mashup Based Method for Event Detection from Social Media

Some events, such as terrorism attacks, earthquakes, and other events that represent tipping points, remain engraved in our memories. Today, through social media, researchers attempt to propose approaches for event detection. However, they are confronted to certain challenges owing to the noise of data propagated throughout social media. In this paper, a new mashup based method for event detection from social media is proposed using hadoop framework. The suggested approach aims at detecting real-world events by exploiting data collected from different social media sites. Indeed, the detected events are characterized by such descriptive dimensions as topic, time and location. Moreover, our approach assures a bilingual event detection. In fact, the proposed approach is able to detect events in English and French languages. In addition, our approach provides a mashup based multidimensional visualization by combining different multimedia components so as to add more details to the detected events. Furthermore, in order to overcome the problems occurring from the processing of big data, we integrated our approach into the hadoop distributed system.

Deep Bayesian network architecture for Big Data mining

Classical Datamining methods are facing various challenges in the era of Big Data. Between the need of fast knowledge extraction and the high flows of data acquired in small slots of time, these methods became shifted. The variability and the veracity of the Big Data perplex the Machine Learning process. The high volume of Big Data yields to a congested learning because the classic methods are designed for small sets of features. Deep Learning has recently emerged in the aim of handling voluminous data. The concept of the Deep induces the conversion of the features into a new abstracted representation in order to optimize an objective. Although the Deep Learning methods are experimentally promising, their parameterization is exhaustive and empirical. To tackle these problems, we utilize the causality and the uncertainty of the Bayesian Network in order to propose a new Deep Bayesian Network architecture. We provide a new learning algorithm for this multi‐layered Bayesian Network with latent variables. We evaluate the proposed architecture and learning algorithms over benchmark datasets. We used high‐dimensional data in order to simulate the Big Data challenges, which are imposed by the volume and veracity aspects. We demonstrate the effectiveness of our contribution under these constraints.