Knowledge-based Features Research Articles

Feature engineering and machine learning for computer-assisted screening of children with speech disorders.

Auditory perceptual analysis (APA) is the main method for clinical assessment of speech-language deficits, which are one of the most prevalent childhood disabilities. However, results from APA are susceptible to intra- and inter-rater variabilities. There are also other limitations of manual or hand transcription-based speech disorder diagnostic methods. There is increased interest in developing automated methods that quantify speech patterns for diagnosing speech disorders in children to address these limitations. Landmark (LM) analysis is an approach that characterizes acoustic events occurring due to sufficiently precise articulatory movements. This work investigates the utilization of LMs for automatic speech disorder detection in children. Besides the LM-based features that have been proposed in existing research, we propose a set of novel knowledge-based features that have not been proposed before. A systematic study and comparison of different linear and nonlinear machine learning classification techniques based on the raw features and the proposed features is conducted to assess the effectiveness of the novel features in classifying speech disorder patients from normal speakers.

MACHINE LEARNING-BASED ECONOMIC DEVELOPMENT MAPPING FROM MULTI-SOURCE OPEN GEOSPATIAL DATA

Abstract. Timely and accurate socioeconomic indicators are the prerequisite for smart social governance. For example, the level of economic development and the structure of population are important statistics for regional or national policy-making. However, the collection of these characteristics usually depends on demographic and social surveys, which are time- and labor-intensive. To address these issues, we propose a machine learning-based approach to estimate and map the economic development from multi-source open available geospatial data, including remote sensing imagery and OpenStreetMap road networks. Specifically, we first extract knowledge-based features from different data sources; then the multi-view graphs are constructed through different perspectives of spatial adjacency and feature similarity; and a multi-view graph neural network (MVGNN) model is built on them and trained in a self-supervised learning manner. Then, the handcrafted features and the learned graph representations are combined to estimate the regional economic development indicators via random forest models. Taking China’s county-level gross domestic product (GDP) as an example, extensive experiments have been conducted and the results demonstrate the effectiveness of the proposed method, and the combination of the knowledge-based and learning-based features can significantly outperform baseline methods. Our proposed approach can advance the goal of acquiring timely and accurate socioeconomic variables through widely accessible geospatial data, which has the potential to extend to more social indicators and other geographic regions to support smart governance and policy-making in the future.

Distributed Feature Sets for Document Specific Key-Phrase Extraction

The need for knowledge and the satisfaction of this need is the fastest growing market in the world. Among all types of data, the textual documents are an incredible source of knowledge. In order not to lose track of this rapid development of information diversity, it is necessary to provide schemes that make it possible to filter out and present specific information from this huge amount of textual data. Since then, key-phrase extraction methods have started to gain importance. A key-phrase extraction system provides a selection of relevant data in textual documents based on languages and corresponding extraction rules. To extract the important concepts from the documents, the system should be able to use special features and self-identify properties of the words in the texts and properties of the documents. In this article, we developed and discussed different sets of features which are unrestricted to form, size and organization of the documents, i.e. a novel set of regular, advanced and external knowledge-based features are proposed. To selectively combine the best features from all three sets of features here we deploy the two different automatic feature selection techniques. Four different datasets are used here to evaluate the performance of the individual, combined and best selected features. The dynamic programming-based feature selection approaches significantly improves the performance in contrast to the different feature sets (individual and combined both) and state-of-the-art.

Generalizable Beat-by-Beat Arrhythmia Detection by Using Weakly Supervised Deep Learning.

Beat-by-beat arrhythmia detection in ambulatory electrocardiogram (ECG) monitoring is critical for the evaluation and prognosis of cardiac arrhythmias, however, it is a highly professional demanding and time-consuming task. Current methods for automatic beat-by-beat arrhythmia detection suffer from poor generalization ability due to the lack of large-sample and finely-annotated (labels are given to each beat) ECG data for model training. In this work, we propose a weakly supervised deep learning framework for arrhythmia detection (WSDL-AD), which permits training a fine-grained (beat-by-beat) arrhythmia detector with the use of large amounts of coarsely annotated ECG data (labels are given to each recording) to improve the generalization ability. In this framework, heartbeat classification and recording classification are integrated into a deep neural network for end-to-end training with only recording labels. Several techniques, including knowledge-based features, masked aggregation, and supervised pre-training, are proposed to improve the accuracy and stability of the heartbeat classification under weak supervision. The developed WSDL-AD model is trained for the detection of ventricular ectopic beats (VEB) and supraventricular ectopic beats (SVEB) on five large-sample and coarsely-annotated datasets and the model performance is evaluated on three independent benchmarks according to the recommendations from the Association for the Advancement of Medical Instrumentation (AAMI). The experimental results show that our method improves the F 1 score of supraventricular ectopic beats detection by 8%–290% and the F1 of ventricular ectopic beats detection by 4%–11% on the benchmarks compared with the state-of-the-art methods of supervised learning. It demonstrates that the WSDL-AD framework can leverage the abundant coarsely-labeled data to achieve a better generalization ability than previous methods while retaining fine detection granularity. Therefore, this framework has a great potential to be used in clinical and telehealth applications. The source code is available at https://github.com/sdnjly/WSDL-AD.

Inner Knowledge-based Img2Doc Scheme for Visual Question Answering

Visual Question Answering (VQA) is a research topic of significant interest at the intersection of computer vision and natural language understanding. Recent research indicates that attributes and knowledge can effectively improve performance for both image captioning and VQA. In this article, an inner knowledge-based Img2Doc algorithm for VQA is presented. The inner knowledge is characterized as the inner attribute relationship in visual images. In addition to using an attribute network for inner knowledge-based image representation, VQA scheme is associated with a question-guided Doc2Vec method for question–answering. The attribute network generates inner knowledge-based features for visual images, while a novel question-guided Doc2Vec method aims at converting natural language text to vector features. After the vector features are extracted, they are combined with visual image features into a classifier to provide an answer. Based on our model, the VQA problem is resolved by textual question answering. The experimental results demonstrate that the proposed method achieves superior performance on multiple benchmark datasets.

Leveraging Knowledge-Based Features With Multilevel Attention Mechanisms for Short Arabic Text Classification

With the wide spread of short texts through social media platforms, there has become a growing need for effective methods for short-text classification. However, short-text classification has always been challenging due to the ambiguity and the data sparsity of the short text. A common solution is to enrich the short text with additional semantic features extracted from external knowledge, such as Wikipedia, to help the classifier better decide on the correct class. Most existing works, however, focused on text written in English and benefited from the existence of entity-linking tools based on English-based knowledge bases. When it comes to the Arabic language, the exploitation of external knowledge to support the classification of Arabic short text has not been widely explored. This work presents an approach for the classification of short Arabic text that exploits both the Wikipedia-based features and the attention mechanism for effective classification. First, Wikipedia entities mentioned in the short text are identified. Then, Wikipedia categories associated with the identified entities are retrieved and filtered to retain only the most relevant categories. A deep learning model with multiple attention mechanisms is then used to encode the short text and the associated category set. Finally, the short text and category representations are combined together to be fed into the classification layer. The use of the attentive model with category filtering leads to highlighting the most important features while reducing the effect of improper features. Finally, the proposed model is evaluated by comparing it with several deep learning models.

Network-based protein-protein interaction prediction method maps perturbations of cancer interactome

The perturbations of protein-protein interactions (PPIs) were found to be the main cause of cancer. Previous PPI prediction methods which were trained with non-disease general PPI data were not compatible to map the PPI network in cancer. Therefore, we established a novel cancer specific PPI prediction method dubbed NECARE, which was based on relational graph convolutional network (R-GCN) with knowledge-based features. It achieved the best performance with a Matthews correlation coefficient (MCC) = 0.84±0.03 and an F1 = 91±2% compared with other methods. With NECARE, we mapped the cancer interactome atlas and revealed that the perturbations of PPIs were enriched on 1362 genes, which were named cancer hub genes. Those genes were found to over-represent with mutations occurring at protein-macromolecules binding interfaces. Furthermore, over 56% of cancer treatment-related genes belonged to hub genes and they were significantly related to the prognosis of 32 types of cancers. Finally, by coimmunoprecipitation, we confirmed that the NECARE prediction method was highly reliable with a 90% accuracy. Overall, we provided the novel network-based cancer protein-protein interaction prediction method and mapped the perturbation of cancer interactome. NECARE is available at: https://github.com/JiajunQiu/NECARE.

Phenology-pigment based automated peanut mapping using sentinel-2 images

ABSTRACT Reliable spatiotemporal crop data are vital for sustainable agricultural management. However, efficient algorithms that can be automatically applied to large regions are scarce, especially for cash crops, since it is hard to distinguish their uniqueness merely from temporal profiles of traditional vegetation indices. The efficiency of knowledge-based temporal features and red-edge pigment indices in characterizing crop growth has been reported in the literature, but the potential of combined applications in identifying crops has not been validated yet. This study fills this gap by developing a knowledge-based automated Peanut mapping Algorithm with a combined consideration of crop Phenology and Pigment content variations (PAPP). Peanut crop has earlier and longer flowering stages compared to other crops such as paddy rice and maize. Peanut fields are distinguished with less variations in anthocyanin and chlorophyll as well as higher carotenoid concentrations. Herein, three phenology and pigment-based indicators were proposed for peanut mapping by exploring the concentration and variations of the chlorophyll, anthocyanin and carotenoid indices, respectively. This PAPP algorithm was validated over large regions (around 250 thousand km2 cropland) covering three provinces of Northeast China using Sentinel-2 time-series images. The results reported that there was 8,371 km2 peanut area in Northeast China in 2018, concentrated in the western Jilin and Liaoning provinces. Validation from the 1,102 field survey sites revealed overall accuracies of 94%, with a kappa index of 0.87 and F1 score of 0.91. The PAPP algorithm was not sensitive to thresholding, and a high classification accuracy could be obtained once the threshold of one indicator was roughly defined. The thresholds could be determined based on the proportions of staple crops (i.e. paddy rice and maize) using the historical agricultural statistical data since peanut fields either show the least or largest values in these three proposed indicators. The PAPP algorithm demonstrates the capabilities of automatic peanut mapping over large regions with no requirements of further training and modifications. This study makes contributions to a sustainable agricultural management society given the potential significant role of legume crops in co-delivering food security and adapting to climate change.

PlantMirP2: An Accurate, Fast and Easy-To-Use Program for Plant Pre-miRNA and miRNA Prediction.

MicroRNAs (miRNAs) are a kind of short non-coding ribonucleic acid molecules that can regulate gene expression. The computational identification of plant miRNAs is of great significance to understanding biological functions. In our previous studies, we have put firstly forward and further developed a set of knowledge-based energy features to construct two plant pre-miRNA prediction tools (plantMirP and riceMirP). However, these two tools cannot be used for miRNA prediction from NGS (Next-Generation Sequencing) data. In addition, for further improving the prediction performance and accessibility, plantMirP2 has been developed. Based on the latest dataset, plantMirP2 achieves a promising performance: 0.9968 (Area Under Curve, AUC), 0.9754 (accuracy), 0.9675 (sensitivity) and 0.9876 (specificity). Additionally, the comparisons with other plant pre-miRNA tools show that plantMirP2 performs better. Finally, the webserver and stand-alone version of plantMirP2 are available.

Multilevel clustering approach driven by continuous glucose monitoring data for further classification of type 2 diabetes

IntroductionMining knowledge from continuous glucose monitoring (CGM) data to classify highly heterogeneous patients with type 2 diabetes according to their characteristics remains unaddressed. A refined clustering method that retrieves hidden...

BlinKey

Virtual Reality (VR) has shown promising potentials in many applications, such as e-business, healthcare, and social networking. Rich information regarding user's activities and their online accounts is stored in VR devices. If it is carelessly unattended, then attackers, including insiders, can make use of the stored information to, for example, perform in-app purchases at the legitimate owner's expenses. Current solutions, mostly following schemes designed for general personal devices, have been proved vulnerable to shoulder-surfing attacks due to the sight blocking caused by the headset. Although there have been efforts trying to fill this gap, they either rely on some highly advanced equipment, such as electrodes to read brainwaves, or introduce heavy cognitive load that has users perform a series of cumbersome authentication tasks. Therefore, an authentication method for VR devices that is robust and convenient is in dire need. In this paper, we present the design, implementation, and evaluation of a two-factor user authentication scheme, BlinKey, for VR devices that are equipped with an eye tracker. A user's secret passcode is a set of recorded rhythms when he/she blinks, together with the unique pupil size variation pattern. We call this passcode as a blinkey, which can be jointly characterized by knowledge-based and biometric features. To examine the performances, BlinKey is implemented on an HTC Vive Pro with a Pupil Labs eye tracker. Through extensive experimental evaluations with 52 participants, we show that our scheme can achieve the average EER as low as 4.0% with only 6 training samples. Besides, it is robust against various types of attacks. BlinKey also exhibits satisfactory usability in terms of login attempts, memorability, and impact of user motions. We also carry out questionnaire-based pre-/post-studies. The survey result indicates that BlinKey is well accepted as a user authentication scheme for VR devices.

Breaking the Limits of Transportation Mode Detection: Applying Deep Learning Approach With Knowledge-Based Features

Activity recognition and transportation mode detection are the key research areas for context-aware systems. In smart environments such as cities, buildings, transportation systems etc., ambient intelligence applications watch on human activities in order to increase the quality of transportation, health, traffic and human-oriented services. Literature review shows that existing studies typically deal with recognition of motorized and non-motorized daily activities under the name of human activity recognition and transport mode detection. Only few studies worked on these problems with rich datasets in terms of the number of classes. To the best of our knowledge, for the first time, a study examines nearly the whole motorized/non-motorized transportation modes of people including still, walk, run, climbing upstairs, climbing downstairs, bicycle, motorbike, car, metro, train, high speed rail (HSR), tram and metrobus, excluding ferry and airplane. To outperform existing solutions on such a large dataset, an extended version of the well-known HTC dataset, especially consisting of similar classes, such as train, metro, tram and high speed rail, we propose a combined solution of an Long Short-Term Memory network and Healing algorithm. In our experiments, we first reveal the limits of traditional machine learning algorithms on such number of classes. In addition to that, apart from other studies exploiting deep learning approach, we then examine the potential input types for LSTM network, including raw sensor data, knowledge-based features and features obtained via Auto Encoder. Experimental results show that our proposed idea, feeding knowledge-based features into frames of LSTM network make a remarkable difference and bring out a robust, orientation-independent and generic solution for these well-known problems including activity recognition and transport mode detection. Besides, we examined the hyper-parameters of our deep learning approach and specified the effective parameter set including window size, number of frames, unit size, dropout rate and batch size. Our test results reveal that we could achieve a success rate of 95.5% and outperform the state-of-the art solutions for 12 different transportation modes with the help of our former algorithm, namely Healing, on a totally journey-independent dataset where instances for train, validation and test datasets are selected from different journeys.

Effective DGA-Domain Detection and Classification with TextCNN and Additional Features

Malicious codes, such as advanced persistent threat (APT) attacks, do not operate immediately after infecting the system, but after receiving commands from the attacker’s command and control (C&C) server. The system infected by the malicious code tries to communicate with the C&C server through the IP address or domain address of the C&C server. If the IP address or domain address is hard-coded inside the malicious code, it can analyze the malicious code to obtain the address and block access to the C&C server through security policy. In order to circumvent this address blocking technique, domain generation algorithms are included in the malware to dynamically generate domain addresses. The domain generation algorithm (DGA) generates domains randomly, so it is very difficult to identify and block malicious domains. Therefore, this paper effectively detects and classifies unknown DGA domains. We extract features that are effective for TextCNN-based label prediction, and add additional domain knowledge-based features to improve our model for detecting and classifying DGA-generated malicious domains. The proposed model achieved 99.19% accuracy for DGA classification and 88.77% accuracy for DGA class classification. We expect that the proposed model can be applied to effectively detect and block DGA-generated domains.

Multiple features for clinical relation extraction: A machine learning approach.

Relation extraction aims to discover relational facts about entity mentions from plain texts. In this work, we focus on clinical relation extraction; namely, given a medical record with mentions of drugs and their attributes, we identify relations between these entities. We propose a machine learning model with a novel set of knowledge-based and BioSentVec embedding features. We systematically investigate the impact of these features with standard distance- and word-based features, conducting experiments on two benchmark datasets of clinical texts from MADE 2018 and n2c2 2018 shared tasks. For comparison with the feature-based model, we utilize state-of-the-art models and three BERT-based models, including BioBERT and Clinical BERT. Our results demonstrate that distance and word features provide significant benefits to the classifier. Knowledge-based features improve classification results only for particular types of relations. The sentence embedding feature provides the largest improvement in results, among other explored features on the MADE corpus. The classifier obtains state-of-the-art performance in clinical relation extraction with F-measure of 92.6%, improving F-measure by 3.5% on the MADE corpus.

A general framework and guidelines for benchmarking computational intelligence algorithms applied to forecasting problems derived from an application domain-oriented survey

Benchmarking computational intelligence algorithms provides valuable knowledge for selecting the best or, at least, the proper algorithm for a certain problem. The experimental results of the computational intelligence techniques applications in various domains, as well as the comparative studies that were reported in the literature can be analyzed and synthesized as development strategies for new successful applications of CI algorithms. Starting from an application domain-oriented survey of selected recently reported research work, the paper presents a general benchmarking framework applicable to computational intelligence algorithms and a set of guidelines for the selection of the best or more suitable CI algorithm for solving forecasting problems. Our approach proposes the integration of software and knowledge engineering best practice towards CI benchmarking, being a computational intelligence engineering methodology. The framework uses two knowledge bases, one for the application domain and one for the CI algorithms, providing heuristic knowledge for a more informed and efficient benchmarking, a case base in which solved problems are recorded with their solution and lessons that were learned, and a knowledge-based problem instance features selection. Some examples of how to apply the framework for problems of forecasting in seismology, environmental protection, hydrology and energy are also discussed. We point out that the framework might be implemented as a software tool (e.g. a decision support system) or as a tool suite. The main conclusion of our research work is that the integration of the derived knowledge from an application domain-oriented survey into the general benchmarking framework along with the set of guidelines for best or proper CI algorithms selection can improve significantly the forecasting accuracy and the response time, in case of real time forecasters.

MilRNApredictor: Genome-free prediction of fungi milRNAs by incorporating k-mer scheme and distance-dependent pair potential

MicroRNA-like small RNAs (milRNAs) with length of 21–22 nucleotides are a type of small non-coding RNAs that are firstly found in Neurospora crassa in 2010. Identifying milRNAs of species without genomic information is a difficult problem. Here, knowledge-based energy features are developed to identify milRNAs by tactfully incorporating k-mer scheme and distance-dependent pair potential. Compared with k-mer scheme, features developed here can alleviate the inherent curse of dimensionality in k-scheme once k becomes large. In addition, milRNApredictor built on novel features performs comparably to k-mer scheme, and achieves sensitivity of 74.21%, and specificity of 75.72% based on 10-fold cross-validation. Furthermore, for novel miRNA prediction, there exists high overlap of results from milRNApredictor and state-of-the-art mirnovo. However, milRNApredictor is simpler to use with reduced requirements of input data and dependencies. Taken together, milRNApredictor can be used to de novo identify fungi milRNAs and other very short small RNAs of non-model organisms.

“Are You Playing a Shooter Again?!” Deep Representation Learning for Audio-Based Video Game Genre Recognition

In this paper, we present a novel computer audition task: audio-based video game genre classification. The aim of this study is threefold: 1) to check the feasibility of the proposed task; 2) to introduce a new corpus: The Game Genre by Audio + Multimodal Extracts (G $^{2}$ AME), collected entirely from social multimedia; and 3) to compare the efficacy of various acoustic feature spaces to classify the G $^{2}$ AME corpus into six game genres using a linear support vector machine classifier. For the classification we extract three different feature representations from the game audio files: 1) Knowledge-based acoustic features; 2) Deep Spectrum features; and 3) quantized Deep Spectrum features using Bag-of-Audio-Words. The Deep Spectrum features are a deep-learning-based representation derived from forwarding the visual representations of the audio instances, in particular spectrograms, mel-spectrograms, chromagrams, and their deltas through deep task-independent pretrained CNNs. Specifically, activations of fully connected layers from three common image classification CNNs, GoogLeNet, AlexNet, and VGG16 are used as feature vectors. Results for the six-genre classification problem indicate the suitability of our deep learning approach for this task. Our best method achieves an accuracy of up to 66.9% unweighted average recall using tenfold cross-validation.

Semi-supervised algorithm with knowledge-based features for learner's proles interoperability

Semi-supervised algorithm with knowledge-based features for learner's proles interoperability

Semi-supervised algorithm with knowledge-based features for learner's profiles interoperability

Nowadays, the user can have several profiles found in different adaptive systems relative to various fields. In particular, adaptive e-learning systems respond to a strong need to adapt to each learner their proposed activities based on the data stored in his/her profile (learning-style, interest, etc.). However, each system can have incomplete data as far as the learner is concerned. Hence, the exchange of the learner's profile data is extremely important in order to enhance his/her learning experience. The exchange requires a matching process so as to resolve the large number of a learner's profiles differences whether in syntax, structure or semantics. In this context, we propose a matching process to automatically detect the similarity between the profile elements. The originality of this process resides in the fact that it rests on a new semi-supervised Tri-Training algorithm which significantly improves the state of the art approaches.

Perceiving Boundaries in Unfamiliar Turkish Makam Music

Music segmentation is a widely researched topic within music perception. Even though there is extensive data on the role of surface structure features and music training (e.g., Deliège, 1987) in segmentation, not much is known yet about the influence of implicit knowledge-based features acquired through musical enculturation. The goal of our study was to fill this gap. Makam music-trained musicians, nonmusicians, and Western listeners marked their segmentations online as they listened to mostly 19th century, unfamiliar Turkish makam tunes, all recorded in a Qānūn timbre on MIDI with retained microtonal structure. In addition, two experts segmented the tunes in a free-time setting. We found considerable within- and across-group agreement, as well as good agreement with the expert segmentations. After transforming each participant’s segmentations into “hits” and “false alarms” based on their match or mismatch with expert segmentations, we observed that musicians overlapped significantly more with expert segmentations than do the other two groups. Segmentations in all three groups were strongly driven by mostly local surface features. Overall, our results are more supportive of a universality claim as proposed by the Gestalt school of psychology than an enculturation claim.