Search

The use of back propagation through the time learning rule enabled the supervised training of deep spiking neural networks to process temporal neuromorphic data. However, their performance is still below non-spiking neural networks. Previous work pointed out that one of the main causes is the limited number of neuromorphic data currently available, which are also difficult to generate. With the goal of overcoming this problem, we explore the usage of auxiliary learning as a means of helping spiking neural networks to identify more general features. Tests are performed on neuromorphic DVS-CIFAR10 and DVS128-Gesture datasets. The results indicate that training with auxiliary learning tasks improves their accuracy, albeit slightly. Different scenarios, including manual and automatic combination losses using implicit differentiation, are explored to analyze the usage of auxiliary tasks.

Unemployment, a significant economic and social challenge, triggers repercussions that affect individual workers and companies, generating a national economic impact. Forecasting the unemployment rate becomes essential for policymakers, allowing them to make short-term estimates, assess economic health, and make informed monetary policy decisions. This paper proposes the innovative GA-LSTM method, which fuses an LSTM neural network with a genetic algorithm to address challenges in unemployment prediction. Effective parameter determination in recurrent neural networks is crucial and a well-known challenge. The research uses the LSTM neural network to overcome complexities and nonlinearities in unemployment predictions, complementing it with a genetic algorithm to optimize the parameters. The central objective is to evaluate recurrent neural network models by comparing them with GA-LSTM to identify the most appropriate model for predicting unemployment in Ecuador using monthly data collected by various organizations. The results demonstrate that the hybrid GA-LSTM model outperforms traditional approaches, such as BiLSTM and GRU, on various performance metrics. This finding suggests that the combination of the predictive power of LSTM with the optimization capacity of the genetic algorithm offers a robust and effective solution to address the complexity of predicting unemployment in Ecuador.

Predicting student dropout is a crucial task in online education. Traditionally, each educational entity (institution, university, faculty, department, etc.) creates and uses its own prediction model starting from its own data. However, that approach is not always feasible or advisable and may depend on the availability of data, local infrastructure, and resources. In those cases, there are various machine learning approaches for sharing data and/or models between educational entities, using a classical centralized machine learning approach or other more advanced approaches such as transfer learning or federated learning. In this paper, we used data from three different LMS Moodle servers representing homogeneous different-sized educational entities. We tested the performance of the different machine learning approaches for the problem of predicting student dropout with multiple educational entities involved. We used a deep learning algorithm as a predictive classifier method. Our preliminary findings provide useful information on the benefits and drawbacks of each approach, as well as suggestions for enhancing performance when there are multiple institutions. In our case, repurposed transfer learning, stacked transfer learning, and centralized approaches produced similar or better results than the locally trained models for most of the entities.

Teacher evaluation is presented as an object of study of great interest, where multiple efforts converge to establish models from the association of heterogeneous data from academic actors, one of these is the students' community, who stands out for their contribution with rich data information for the establishment of teacher evaluation in higher education. This study aims to present the search results for references on the prediction of teacher evaluation based on the associated data provided by the performance of university students. For this purpose, a systematic literature review was carried out, established by the phases of planning (search objective, research questions, inclusion and exclusion criteria), search and selection (literature control group and keywords, the definition of the search string, results filtering), and extraction (synthesis of the contributions). As a result, a set of references on the application of predictions is obtained, focused on educational data mining techniques, such as Fuzzy logic, Fuzzy clustering, Fuzzy Neural Network (FNN), Neural networks, multilayer perceptron (MLP), Decision Trees, Logistic Regression, Random Forest Classifier, Naïve Bayes Classifier, Support Vector Machine (SVM), K-Nearest-Neighbor (KNN), and Associative classification model. In conclusion, prediction and mining techniques have been widely explored; however, teacher evaluation is in the process of growth with particular emphasis on fuzzy principles, considering that human decision-making is developed with uncertainty, which is strongly related to human behavior.

Abstract Background Pancreatic ductal adenocarcinoma (PDAC) is a highly lethal cancer, requiring novel treatments to target both cancer cells and cancer stem cells (CSCs). Altered splicing is emerging as both a novel cancer hallmark and an attractive therapeutic target. The core splicing factor SF3B1 is heavily altered in cancer and can be inhibited by Pladienolide-B, but its actionability in PDAC is unknown. We explored the presence and role of SF3B1 in PDAC and interrogated its potential as an actionable target. Methods SF3B1 was analyzed in PDAC tissues, an RNA-seq dataset, and publicly available databases, examining associations with splicing alterations and key features/genes. Functional assays in PDAC cell lines and PDX-derived CSCs served to test Pladienolide-B treatment effects in vitro, and in vivo in zebrafish and mice. Results SF3B1 was overexpressed in human PDAC and associated with tumor grade and lymph-node involvement. SF3B1 levels closely associated with distinct splicing event profiles and expression of key PDAC players (KRAS, TP53). In PDAC cells, Pladienolide-B increased apoptosis and decreased multiple tumor-related features, including cell proliferation, migration, and colony/sphere formation, altering AKT and JNK signaling, and favoring proapoptotic splicing variants (BCL-XS/BCL-XL, KRASa/KRAS, Δ133TP53/TP53). Importantly, Pladienolide-B similarly impaired CSCs, reducing their stemness capacity and increasing their sensitivity to chemotherapy. Pladienolide-B also reduced PDAC/CSCs xenograft tumor growth in vivo in zebrafish and in mice. Conclusion SF3B1 overexpression represents a therapeutic vulnerability in PDAC, as altered splicing can be targeted with Pladienolide-B both in cancer cells and CSCs, paving the way for novel therapies for this lethal cancer.

Abstract Background Associative Classification, a combination of two important and different fields (classification and association rule mining), aims at building accurate and interpretable classifiers by means of association rules. A major problem in this field is that existing proposals do not scale well when Big Data are considered. In this regard, the aim of this work is to propose adaptations of well-known associative classification algorithms (CBA and CPAR) by considering different Big Data platforms (Spark and Flink). Results An experimental study has been performed on 40 datasets (30 classical datasets and 10 Big Data datasets). Classical data have been used to find which algorithms perform better sequentially. Big Data dataset have been used to prove the scalability of Big Data proposals. Results have been analyzed by means of non-parametric tests. Results proved that CBA-Spark and CBA-Flink obtained interpretable classifiers but it was more time consuming than CPAR-Spark or CPAR-Flink. In this study, it was demonstrated that the proposals were able to run on Big Data (file sizes up to 200 GBytes). The analysis of different quality metrics revealed that no statistical difference can be found for these two approaches. Finally, three different metrics (speed-up, scale-up and size-up) have also been analyzed to demonstrate that the proposals scale really well on Big Data. Conclusions The experimental study has revealed that sequential algorithms cannot be used on large quantities of data and approaches such as CBA-Spark, CBA-Flink, CPAR-Spark or CPAR-Flink are required. CBA has proved to be very useful when the main goal is to obtain highly interpretable results. However, when the runtime has to be minimized CPAR should be used. No statistical difference could be found between the two proposals in terms of quality of the results except for the interpretability of the final classifiers, CBA being statistically better than CPAR.

In this paper we present a Competitive Rate-Based Algorithm (CRBA) that approximates operation of a Competitive Spiking Neural Network (CSNN). CRBA is based on modeling of the competition between neurons during a sample presentation, which can be reduced to ranking of the neurons based on a dot product operation and the use of a discrete Expectation Maximization algorithm; the latter is equivalent to the spike time-dependent plasticity rule. CRBA's performance is compared with that of CSNN on the MNIST and Fashion-MNIST datasets. The results show that CRBA performs on par with CSNN, while using three orders of magnitude less computational time. Importantly, we show that the weights and firing thresholds learned by CRBA can be used to initialize CSNN's parameters that results in its much more efficient operation.

Multi-view learning analyzes the information from several perspectives and has largely been applied on semi-supervised contexts. It has not been extensively analyzed for inducing interpretable rule-based classifiers. We present a multi-view and grammar-based genetic programming model for inducing rules for semi-supervised contexts. It evolves several populations and views, and promotes both accuracy and agreement among the views. This work details how and why common practices may not produce the expected results when inducing rule-based classifiers under this methodology.

Background: Dysregulation of splicing variants (SVs) expression has recently emerged as a novel cancer hallmark. Although the generation of aberrant SVs (e.g. AR-v7/sst5TMD4/etc.) is associated to prostate-cancer (PCa) aggressiveness and/or castration-resistant PCa (CRPC) development, whether the molecular reason behind such phenomena might be linked to a dysregulation of the cellular machinery responsible for the splicing process [spliceosome-components (SCs) and splicing-factors (SFs)] has not been yet explored. Methods: Expression levels of 43 key SCs and SFs were measured in two cohorts of PCa-samples: 1) Clinically-localized formalin-fixed paraffin-embedded PCa-samples (n = 84), and 2) highly-aggressive freshly-obtained PCa-samples (n = 42). Findings: A profound dysregulation in the expression of multiple components of the splicing machinery (i.e. 7 SCs/19 SFs) were found in PCa compared to their non-tumor adjacent-regions. Notably, overexpression of SNRNP200, SRSF3 and SRRM1 (mRNA and/or protein) were associated with relevant clinical (e.g. Gleason score, T-Stage, metastasis, biochemical recurrence, etc.) and molecular (e.g. AR-v7 expression) parameters of aggressiveness in PCa-samples. Functional (cell-proliferation/migration) and mechanistic [gene-expression (qPCR) and protein-levels (western-blot)] assays were performed in normal prostate cells (PNT2) and PCa-cells (LNCaP/22Rv1/PC-3/DU145 cell-lines) in response to SNRNP200, SRSF3 and/or SRRM1 silencing (using specific siRNAs) revealed an overall decrease in proliferation/migration-rate in PCa-cells through the modulation of key oncogenic SVs expression levels (e.g. AR-v7/PKM2/XBP1s) and alteration of oncogenic signaling pathways (e.g. p-AKT/p-JNK). Interpretation: These results demonstrate that the spliceosome is drastically altered in PCa wherein SNRNP200, SRSF3 and SRRM1 could represent attractive novel diagnostic/prognostic and therapeutic targets for PCa and CRPC. Keywords: Prostate cancer, Splicing, Spliceosome, SNRNP200, SRSF3, SRRM1, Therapeutic target

Skin cancer is one of the most common types of cancers in the world, with melanoma being the most lethal form. Automatic melanoma diagnosis from skin images has recently gained attention within the machine learning community, due to the complexity involved. In the past few years, convolutional neural network models have been commonly used to approach this issue. This type of model, however, presents disadvantages that sometimes hamper its application in real-world situations, e.g., the construction of transformation-invariant models and their inability to consider spatial hierarchies between entities within an image. Recently, Dynamic Routing between Capsules architecture (CapsNet) has been proposed to overcome such limitations. This work is aimed at proposing a new architecture which combines convolutional blocks with a customized CapsNet architecture, allowing for the extraction of richer abstract features. This architecture uses high-quality 299×299×3 skin lesion images, and a hyper-tuning of the main parameters is performed in order to ensure effective learning under limited training data. An extensive experimental study on eleven image datasets was conducted where the proposal significantly outperformed several state-of-the-art models. Finally, predictions made by the model were validated through the application of two modern model-agnostic interpretation tools.

The transition from high school to university is a critical step and many students head toward failure just because their final degree option was not the right choice. Both students’ preferences and skills play an important role in choosing the degree that best fits them, so an analysis of these attitudes during the high school can minimize the drop out in a posteriori learning period like university. We propose a subgroup discovery algorithm based on grammars to extract itemsets and relationships that represent any type of homogeneity and regularity in data from a supervised context. This supervised context is cornerstone, considering a single item or a set of them as interesting and distinctive. The proposed algorithm supports the students’ final degree decision by extracting relations among different students’ skills and preferences during the high school period. The idea is to be able to provide advices with regard to what is the best degree option for each specific skill and student. In this regard, the use of grammars is essential since it enables subjective and external knowledge to be included during the mining process. The proposed algorithm has been compared against different subgroup discovery algorithms, achieving excellent results. A real-world experimental analysis has been developed at King Abdulaziz University, one of the most important universities in Saudi Arabia, where there is a special interest in introducing models to understand the students’ skills to guide them accordingly.

In recent years, evolutionary algorithms have been used for classification tasks. However, only a limited number of comparisons exist between classification genetic rule-based systems and gene expression programming rule-based systems. In this paper, a new algorithm for classification using gene expression programming is proposed to accomplish this task, which was compared with several classical state-of-the-art rule-based classifiers. The proposed classifier uses a Michigan approach; the evolutionary process with elitism is guided by a token competition that improves the exploration of fitness surface. Individuals that cover instances, covered previously by others individuals, are penalized. The fitness function is constructed by the multiplying three factors: sensibility, specificity and simplicity. The classifier was constructed as a decision list, sorted by the positive predictive value. The most numerous class was used as the default class. Until now, only numerical attributes are allowed and a mono objective algorithm that combines the three fitness factors is implemented. Experiments with twenty benchmark data sets have shown that our approach is significantly better in validation accuracy than some genetic rule-based state-of-the-art algorithms (i.e., SLAVE, HIDER, Tan, Falco, Bojarczuk and CORE) and not significantly worse than other better algorithms (i.e., GASSIST, LOGIT-BOOST and UCS).

Recommending classification algorithms is an open research problem the solution to which is of tremendous value for practitioners and non-experts data mining users such as educators. This paper proposes a new meta-learning framework for educational domains based on the use of multi-label learning for selecting the best classification algorithms in order to predict students’ performance. In short, the framework considers an offline phase where statistical tests are performed to find the subset of algorithms that achieves the best performance over the repository of educational data sets. The subset of algorithms along with the meta-features extracted from the training data are used to generate a multi-label data set. A multi-label classifier is then trained and, in an online phase, this model is used to recommend the most suitable classification algorithms to be applied to new unseen data sets. This new multi-label meta-learning approach has been applied to a repository of educational data sets generated from Moodle usage data. The results obtained show significant improvement compared with a previous nearest neighbor proposal, demonstrating the suitability of the new framework.

This paper shows how to use evolutionary algorithms to data mining methods for discovering prediction rules in databases. These rules will be used to improve web-based adpative hypermedia courses. The idea is to discover important rules among the usage data picked up during the students’ execuctions. This information may be very useful to the author of the course, who can decide what modification will be the most apropiate to improve the performance of students. In order to do the discovering of rules we have used grammar-based genetic programming (GBGP) with multiobjective optimization technics. In this work we also present a graphic tools for discovering rules to facilitate the usage of the proposed methodology to improve the courses.

En el este artículo vamos a presentar una sistema para la construcción de curso hipermedia adaptativos basados en web. Cualquier profesor utilizando nuestro sistema puede crear rápidamente un curso basado en web que además permite la adaptación tanto de la presentación como de la navegación dependiendo del nivel de conocimiento de cada alumno en particular. Para desarrollarlo hemos utilizado de base el sistema adaptativo genérico AHA (de Bra, 98), al que le hemos realizado varias modificaciones importantes para especializarlo y potenciarlo en el área de educación. Como ejemplo de utilización vamos a mostrar un curso de linux que hemos construido utilizando nuestro sistema y que actualmente está siendo utilizado por estudiantes de la Universidad de Córdoba. También estamos desarrollando una herramienta autor para que el profesor pueda construir el curso fácilmente sin la necesidad de conocer la estructura interna de AHA.