Shallow Learning Algorithms Research Articles

Combining categorical boosting and Shapley additive explanations for building an interpretable ensemble classifier for identifying mineralization-related geochemical anomalies

The vast majority of shallow and deep learning techniques used to identify mineralization-related geochemical anomalies are black-box algorithms that lack the ability to elucidate the individual contributions of each element towards the model predictions. In addition, most of the anomaly identification models established by both shallow and deep learning algorithms lack robustness. Establishing interpretable and robust machine learning models is a challenge in applying machine learning techniques to geochemical anomaly identification. To this end, the categorical boosting (CatBoost) algorithm was employed to build a robust ensemble classifier to identify mineralization-related anomalies from the 1:50,000 geochemical reconnaissance data (stream sediment survey) in the Yeniugou area of Xinjiang (China). The receiver operating characteristic curve (ROC) and precision-recall (P-R) curve of the ensemble model were plotted, and the area under the ROC curve (AUC) as well as the area under the P-R curve (AUPRC) of the ensemble model were calculated to measure the performance of the ensemble model. The ROC curve of the ensemble model approximates that of the perfect classification model. The P-R curve of the ensemble model is close to the upper right corner of the P-R space. The AUC and AUPRC values of the ensemble model reaches 0.9981 and 0.7816, respectively. The identified polymetallic mineralization-related geochemical anomalies account for 3% of the whole exploration area, correctly identifying all known polymetallic deposits. To enhance the interpretability of the CatBoost model, the Shapley additive explanations (SHAP) tool was adopted to graphically interpret the predictions of the ensemble model. The graphic interpretation shows that the importance order of the 14 elements is Ni-Au-Ag-Sn-As-Cr-Zn-Cu-Pb-Sb-W-Bi-Mo-Co. Cu and Ni are most likely metallogenic elements of the study area. Cu interacts with Ni, Ag, As, Sn, Cr, Zn, Pb, Sb, W, Bi, and Co; and Ni interacts with Au, Sn, As, Zn, Cu, W, Bi, and Co. Two polymetallic prospective areas were delineated in the study area. One is Cu-Ni-polymetallic mineralization prospective area, and the other is Ni-polymetallic mineralization prospective area. It can be concluded that the combination of CatBoost and SHAP is an effective way to construct an interpretable ensemble model with high-performance and robustness in identifying mineralization-related geochemical anomalies.

PredPSP: a novel computational tool to discover pathway-specific photosynthetic proteins in plants.

Photosynthetic proteins play a crucial role in agricultural productivity by harnessing light energy for plant growth. Understanding these proteins, especially within C3 and C4 pathways, holds promise for improving crops in challenging environments. Despite existing models, a comprehensive computational framework specifically targeting plant photosynthetic proteins is lacking. The underutilization of plant datasets in computational algorithms accentuates the gap this study aims to fill by introducing a novel sequence-based computational method for identifying these proteins. The scope of this study encompassed diverse plant species, ensuring comprehensive representation across C3 and C4 pathways. Utilizing six deep learning models and seven shallow learning algorithms, paired with six sequence-derived feature sets followed by feature selection strategy, this study developed a comprehensive model for prediction of plant-specific photosynthetic proteins. Following 5-fold cross-validation analysis, LightGBM with 65 and 90 LGBM-VIM selected features respectively emerged as the best models for C3 (auROC: 91.78%, auPRC: 92.55%) and C4 (auROC: 99.05%, auPRC: 99.18%) plants. Validation using an independent dataset confirmed the robustness of the proposed model for both C3 (auROC: 87.23%, auPRC: 88.40%) and C4 (auROC: 92.83%, auPRC: 92.29%) categories. Comparison with existing methods demonstrated the superiority of the proposed model in predicting plant-specific photosynthetic proteins. This study further established a free online prediction server PredPSP ( https://iasri-sg.icar.gov.in/predpsp/ ) to facilitate ongoing efforts for identifying photosynthetic proteins in C3 and C4 plants. Being first of its kind, this study offers valuable insights into predicting plant-specific photosynthetic proteins which holds significant implications for plant biology.

ScMGATGRN: a multiview graph attention network-based method for inferring gene regulatory networks from single-cell transcriptomic data.

The gene regulatory network (GRN) plays a vital role in understanding the structure and dynamics of cellular systems, revealing complex regulatory relationships, and exploring disease mechanisms. Recently, deep learning (DL)-based methods have been proposed to infer GRNs from single-cell transcriptomic data and achieved impressive performance. However, these methods do not fully utilize graph topological information and high-order neighbor information from multiple receptive fields. To overcome those limitations, we propose a novel model based on multiview graph attention network, namely, scMGATGRN, to infer GRNs. scMGATGRN mainly consists of GAT, multiview, and view-level attention mechanism. GAT can extract essential features of the gene regulatory network. The multiview model can simultaneously utilize local feature information and high-order neighbor feature information of nodes in the gene regulatory network. The view-level attention mechanism dynamically adjusts the relative importance of node embedding representations and efficiently aggregates node embedding representations from two views. To verify the effectiveness of scMGATGRN, we compared its performance with 10 methods (five shallow learning algorithms and five state-of-the-art DL-based methods) on seven benchmark single-cell RNA sequencing (scRNA-seq) datasets from five cell lines (two in human and three in mouse) with four different kinds of ground-truth networks. The experimental results not only show that scMGATGRN outperforms competing methods but also demonstrate the potential of this model in inferring GRNs. The code and data of scMGATGRN are made freely available on GitHub (https://github.com/nathanyl/scMGATGRN).

ICRBP-LKHA: Large convolutional kernel and hybrid channel-spatial attention for identifying circRNA-RBP interaction sites.

Circular RNAs (circRNAs) play vital roles in transcription and translation. Identification of circRNA-RBP (RNA-binding protein) interaction sites has become a fundamental step in molecular and cell biology. Deep learning (DL)-based methods have been proposed to predict circRNA-RBP interaction sites and achieved impressive identification performance. However, those methods cannot effectively capture long-distance dependencies, and cannot effectively utilize the interaction information of multiple features. To overcome those limitations, we propose a DL-based model iCRBP-LKHA using deep hybrid networks for identifying circRNA-RBP interaction sites. iCRBP-LKHA adopts five encoding schemes. Meanwhile, the neural network architecture, which consists of large kernel convolutional neural network (LKCNN), convolutional block attention module with one-dimensional convolution (CBAM-1D) and bidirectional gating recurrent unit (BiGRU), can explore local information, global context information and multiple features interaction information automatically. To verify the effectiveness of iCRBP-LKHA, we compared its performance with shallow learning algorithms on 37 circRNAs datasets and 37 circRNAs stringent datasets. And we compared its performance with state-of-the-art DL-based methods on 37 circRNAs datasets, 37 circRNAs stringent datasets and 31 linear RNAs datasets. The experimental results not only show that iCRBP-LKHA outperforms other competing methods, but also demonstrate the potential of this model in identifying other RNA-RBP interaction sites.

Comparative Analysis of Machine Learning and Deep Learning Algorithms for Assessing Agricultural Product Quality Using NIRS.

The success of near-infrared spectroscopy (NIRS) analysis hinges on the precision and robustness of the calibration model. Shallow learning (SL) algorithms like partial least squares discriminant analysis (PLS-DA) often fall short in capturing the interrelationships between adjacent spectral variables, and the analysis results are easily affected by spectral noise, which dramatically limits the breadth and depth of applications of NIRS. Deep learning (DL) methods, with their capacity to discern intricate features from limited samples, have been progressively integrated into NIRS. In this paper, two discriminant analysis problems, including wheat kernels and Yali pears as examples, and several representative calibration models were used to research the robustness and effectiveness of the model. Additionally, this article proposed a near-infrared calibration model, which was based on the Gramian angular difference field method and coordinate attention convolutional neural networks (G-CACNNs). The research results show that, compared with SL, spectral preprocessing has a smaller impact on the analysis accuracy of consensus learning (CL) and DL, and the latter has the highest analysis accuracy in the modeling results using the original spectrum. The accuracy of G-CACNNs in two discrimination tasks was 98.48% and 99.39%. Finally, this research compared the performance of various models under noise to evaluate the robustness and noise resistance of the proposed method.

ProkDBP: Toward more precise identification of prokaryotic DNA binding proteins.

Prokaryotic DNA binding proteins (DBPs) play pivotal roles in governing gene regulation, DNA replication, and various cellular functions. Accurate computational models for predicting prokaryotic DBPs hold immense promise in accelerating the discovery of novel proteins, fostering a deeper understanding of prokaryotic biology, and facilitating the development of therapeutics targeting for potential disease interventions. However, existing generic prediction models often exhibit lower accuracy in predicting prokaryotic DBPs. To address this gap, we introduce ProkDBP, a novel machine learning-driven computational model for prediction of prokaryotic DBPs. For prediction, a total of nine shallow learning algorithms and five deep learning models were utilized, with the shallow learning models demonstrating higher performance metrics compared to their deep learning counterparts. The light gradient boosting machine (LGBM), coupled with evolutionarily significant features selected via random forest variable importance measure (RF-VIM) yielded the highest five-fold cross-validation accuracy. The model achieved the highest auROC (0.9534) and auPRC (0.9575) among the 14 machine learning models evaluated. Additionally, ProkDBP demonstrated substantial performance with an independent dataset, exhibiting higher values of auROC (0.9332) and auPRC (0.9371). Notably, when benchmarked against several cutting-edge existing models, ProkDBP showcased superior predictive accuracy. Furthermore, to promote accessibility and usability, ProkDBP (https://iasri-sg.icar.gov.in/prokdbp/) is available as an online prediction tool, enabling free access to interested users. This tool stands as a significant contribution, enhancing the repertoire of resources for accurate and efficient prediction of prokaryotic DBPs.

Resolution of tonic concentrations of highly similar neurotransmitters using voltammetry and deep learning

With advances in our understanding regarding the neurochemical underpinnings of neurological and psychiatric diseases, there is an increased demand for advanced computational methods for neurochemical analysis. Despite having a variety of techniques for measuring tonic extracellular concentrations of neurotransmitters, including voltammetry, enzyme-based sensors, amperometry, and in vivo microdialysis, there is currently no means to resolve concentrations of structurally similar neurotransmitters from mixtures in the in vivo environment with high spatiotemporal resolution and limited tissue damage. Since a variety of research and clinical investigations involve brain regions containing electrochemically similar monoamines, such as dopamine and norepinephrine, developing a model to resolve the respective contributions of these neurotransmitters is of vital importance. Here we have developed a deep learning network, DiscrimNet, a convolutional autoencoder capable of accurately predicting individual tonic concentrations of dopamine, norepinephrine, and serotonin from both in vitro mixtures and the in vivo environment in anesthetized rats, measured using voltammetry. The architecture of DiscrimNet is described, and its ability to accurately predict in vitro and unseen in vivo concentrations is shown to vastly outperform a variety of shallow learning algorithms previously used for neurotransmitter discrimination. DiscrimNet is shown to generalize well to data captured from electrodes unseen during model training, eliminating the need to retrain the model for each new electrode. DiscrimNet is also shown to accurately predict the expected changes in dopamine and serotonin after cocaine and oxycodone administration in anesthetized rats in vivo. DiscrimNet therefore offers an exciting new method for real-time resolution of in vivo voltammetric signals into component neurotransmitters.

A Comprehensive Discourse on Shallow Learning and its Applications

Shallow learning, a fundamental approach in machine learning, encompasses a variety of algorithms and techniques aimed at learning patterns and making predictions from labelled data. Unlike deep learning, which involves complex architectures with multiple layers of abstraction, shallow learning focuses on simpler models with limited complexity. This abstract explores the essence of shallow learning, its algorithms, applications, and challenges. Shallow learning algorithms include classic methods such as decision trees, support vector machines, k-nearest neighbours, and logistic regression, among others. These algorithms are typically trained using supervised learning techniques, where the model learns from input-output pairs to make predictions on new, unseen data. Shallow learning models excel in tasks such as classification and regression, where the goal is to assign labels or predict continuous values to input data. Applications of shallow learning span across various domains, including healthcare, finance, marketing, and cyber security. In healthcare, shallow learning models are used for disease diagnosis and prognosis prediction based on patient data. In finance, these models aid in fraud detection, credit scoring, and stock market prediction. Marketing applications involve customer segmentation and churn prediction, while in cyber security; shallow learning is utilized for malware detection and network intrusion detection.

Supervised learning for automatic emotion recognition in Parkinson’s disease through smartwatch signals

People with Parkinson’s Disease (PwPD) usually experience several neuropsychiatric signs such as anxiety, depression, and negative emotions that contribute to disability and worsening of quality of life. Notwithstanding, the assessment of these symptoms are largely underrated, subjective and difficult due to a large overlapping with other PD symptoms, like hypomimia and bradikinesia. The aim and novelty of the current work is to study and validate a method for automatic emotion recognition in PwPD during daily living through autonomic signals acquired by acceptable and low-cost consumer technology. The best shallow learning algorithm and the best minimal feature set are individuated. 11 PwPD and 8 subjects with no history of neurological injury or illness were enrolled in the study. Participants were asked to watch video clips purposely selected to arouse emotions, and annotate arousal and valence of emotions triggered by video clips, while their heart rate, skin conductance, and temperature were recorded by a smartwatch. Smartwatch data was used for features extraction, while participants’ reported arousal and valence were used as gold-standard to train machine learning algorithms for emotion classification (low/high arousal, positive/negative valence). Different feature sets and different algorithms (i.e. decision tree (DT), random forest (RF), support vector machine (SVM), and multilayer perceptron (MLP)) were evaluated to find the best solution for each group of participant. In each group of participants, it was possible to find a combination of feature set and algorithm to reach a classification accuracy greater than 90%. The Random Forest reached the best performance in both groups and for both valence and arousal. For each classification task (valence or arousal, PwPD or controls), the best model was selected and the minimal feature set was found by performing a recursive feature elimination based on the Shapley value. A lower accuracy of appraisal emerged for arousal compared to valence. Obtained results showed the feasibility of automatic emotion recognition in PwPDs through autonomic signals. Autonomic dysfunction in PwPDs may explain the lower arousal accuracy. The findings warrant confirmation from trials on larger samples and there are open issues to be deepened in future work.

A robust intrusion detection system based on a shallow learning model and feature extraction techniques.

The escalating prevalence of cybersecurity risks calls for a focused strategy in order to attain efficient resolutions. This study introduces a detection model that employs a tailored methodology integrating feature selection using SHAP values, a shallow learning algorithm called PV-DM, and machine learning classifiers like XGBOOST. The efficacy of our suggested methodology is highlighted by employing the NSL-KDD and UNSW-NB15 datasets. Our approach in the NSL-KDD dataset exhibits exceptional performance, with an accuracy of 98.92%, precision of 98.92%, recall of 95.44%, and an F1-score of 96.77%. Notably, this performance is achieved by utilizing only four characteristics, indicating the efficiency of our approach. The proposed methodology achieves an accuracy of 82.86%, precision of 84.07%, recall of 77.70%, and an F1-score of 80.20% in the UNSW-NB15 dataset, using only six features. Our research findings provide substantial evidence of the enhanced performance of the proposed model compared to a traditional deep-learning model across all performance metrics.

DiamondNet: A Neural-Network-Based Heterogeneous Sensor Attentive Fusion for Human Activity Recognition.

With the proliferation of intelligent sensors integrated into mobile devices, fine-grained human activity recognition (HAR) based on lightweight sensors has emerged as a useful tool for personalized applications. Although shallow and deep learning algorithms have been proposed for HAR problems in the past decades, these methods have limited capability to exploit semantic features from multiple sensor types. To address this limitation, we propose a novel HAR framework, DiamondNet, which can create heterogeneous multisensor modalities, denoise, extract, and fuse features from a fresh perspective. In DiamondNet, we leverage multiple 1-D convolutional denoising autoencoders (1-D-CDAEs) to extract robust encoder features. We further introduce an attention-based graph convolutional network to construct new heterogeneous multisensor modalities, which adaptively exploit the potential relationship between different sensors. Moreover, the proposed attentive fusion subnet, which jointly employs a global-attention mechanism and shallow features, effectively calibrates different-level features of multiple sensor modalities. This approach amplifies informative features and provides a comprehensive and robust perception for HAR. The efficacy of the DiamondNet framework is validated on three public datasets. The experimental results demonstrate that our proposed DiamondNet outperforms other state-of-the-art baselines, achieving remarkable and consistent accuracy improvements. Overall, our work introduces a new perspective on HAR, leveraging the power of multiple sensor modalities and attention mechanisms to significantly improve the performance.

A Novel Semantic Segmentation Approach Using Improved SegNet and DSC in Remote Sensing Images

An improved SegNet semantic segmentation model is proposed to address the issue of traditional classification algorithms and shallow learning algorithms not being suitable for extracting information from high-resolution remote sensing images. During the research process, space remote sensing images obtained from the GF-1 satellite were used as the data source. In order to improve the operational efficiency of the encoding network, the pooling layer in the encoding network is removed and the ordinary convolutional layer is replaced with a depth-wise separable convolution. By decoding the last layer of the network to obtain the reshaped output results, and then calculating the probability of each classification using a Softmax classifier, the classification of pixels can be achieved. The output result of the classifier is the final result of the remote sensing image semantic segmentation model. The results showed that the proposed algorithm had the highest Kappa coefficient of 0.9531, indicating good classification performance.

Machine learning-based evaluation of eccentricity and acoustic impedance in oil well using VDL data

Logging procedure and interpretation are complicated mainly because of the large amount of different data inherited. One way to facilitate the problem is to use fewer tools but keep the same amount of information which can be feasible with the assistance of machine learning models. In this paper, Variable Density Logging data was used to evaluate the eccentricity and the acoustic impedance average of the medium behind the casing of oil wells. Different shallow and deep learning architectures were trained, and a Principal Components Analysis feature extraction approach was investigated. Actual data from acoustic logging tools acquired in different wells were adopted. A supervised learning approach was explored, and the K-Nearest Neighbors model obtained the best results when accounting for the cost-benefit and complexity of the models. State-of-the-art hyperparameter optimization techniques were used to tune the non-trainable parameters of the shallow learning algorithms evaluated. The proposed procedure was able to achieve mean balanced adjacency accuracy results greater than 80% for the free pipe pass and repeated pass of a similar well and results greater than 70% for the main pass of a distinguished oil well; which is a more challenging scenario.

DBPMod: a supervised learning model for computational recognition of DNA-binding proteins in model organisms.

DNA-binding proteins (DBPs) play critical roles in many biological processes, including gene expression, DNA replication, recombination and repair. Understanding the molecular mechanisms underlying these processes depends on the precise identification of DBPs. In recent times, several computational methods have been developed to identify DBPs. However, because of the generic nature of the models, these models are unable to identify species-specific DBPs with higher accuracy. Therefore, a species-specific computational model is needed to predict species-specific DBPs. In this paper, we introduce the computational DBPMod method, which makes use of a machine learning approach to identify species-specific DBPs. For prediction, both shallow learning algorithms and deep learning models were used, with shallow learning models achieving higher accuracy. Additionally, the evolutionary features outperformed sequence-derived features in terms of accuracy. Five model organisms, including Caenorhabditis elegans, Drosophila melanogaster, Escherichia coli, Homo sapiens and Mus musculus, were used to assess the performance of DBPMod. Five-fold cross-validation and independent test set analyses were used to evaluate the prediction accuracy in terms of area under receiver operating characteristic curve (auROC) and area under precision-recall curve (auPRC), which was found to be ~89-92% and ~89-95%, respectively. The comparative results demonstrate that the DBPMod outperforms 12 current state-of-the-art computational approaches in identifying the DBPs for all five model organisms. We further developed the web server of DBPMod to make it easier for researchers to detect DBPs and is publicly available at https://iasri-sg.icar.gov.in/dbpmod/. DBPMod is expected to be an invaluable tool for discovering DBPs, supplementing the current experimental and computational methods.

The use of the general thermal sensation discriminant model based on CNN for room temperature regulation by online brain-computer interface

Brain-computer interface (BCI) technology can realize dynamic room temperature adjustment based on individual real-time thermal sensation, which can provide the basis for future intelligent buildings. However, the generalization ability of previous thermal sensation discrimination model (TSDM) is limited, which is a serious obstacle to the application. In this paper, a general TSDM was developed by using convolutional neural network (CNN), which can be well applied to new subjects.In the study, the CNN-TSDM was established and evaluated based on the offline experimental data, and then the BCI closed-loop online room temperature control experiment was carried out based on this CNN-TSDM to further verify. The offline analysis results show that the recognition performance of CNN-TSDM in new subjects is significantly higher than that of typical shallow learning algorithms, and its area under the ROC curve (AUC) value reaches 0.789. In the online experiments of the two simulated environments, BCI using the CNN-TSDM dynamically controlled the air conditioning to improve the room temperature to the comfortable level according to the subjects' thermal sensation. The subjective score of subjects decreased from 3.1 to 3.0 for the hot uncomfortable to 1.1 and 1.2 for the cool comfortable (p < 0.001, p < 0.001). Moreover, in a hotter simulated experimental environment, BCI automatically controlled the air conditioner for longer cooling to obtain a same degree of thermal comfort. The total cooling time (p < 0.05) and the single cooling time (p < 0.05) of the air conditioner were significantly increased. This further confirmed the effectiveness and robustness of the general CNN-TSDM.

RBPLight: a computational tool for discovery of plant-specific RNA-binding proteins using light gradient boosting machine and ensemble of evolutionary features.

RNA-binding proteins (RBPs) are essential for post-transcriptional gene regulation in eukaryotes, including splicing control, mRNA transport and decay. Thus, accurate identification of RBPs is important to understand gene expression and regulation of cell state. In order to detect RBPs, a number of computational models have been developed. These methods made use of datasets from several eukaryotic species, specifically from mice and humans. Although some models have been tested on Arabidopsis, these techniques fall short of correctly identifying RBPs for other plant species. Therefore, the development of a powerful computational model for identifying plant-specific RBPs is needed. In this study, we presented a novel computational model for locating RBPs in plants. Five deep learning models and ten shallow learning algorithms were utilized for prediction with 20 sequence-derived and 20 evolutionary feature sets. The highest repeated five-fold cross-validation accuracy, 91.24% AU-ROC and 91.91% AU-PRC, was achieved by light gradient boosting machine. While evaluated using an independent dataset, the developed approach achieved 94.00% AU-ROC and 94.50% AU-PRC. The proposed model achieved significantly higher accuracy for predicting plant-specific RBPs as compared to the currently available state-of-art RBP prediction models. Despite the fact that certain models have already been trained and assessed on the model organism Arabidopsis, this is the first comprehensive computer model for the discovery of plant-specific RBPs. The web server RBPLight was also developed, which is publicly accessible at https://iasri-sg.icar.gov.in/rbplight/, for the convenience of researchers to identify RBPs in plants.

Do we really need a new theory to understand over-parameterization?

This century saw an unprecedented increase of public and private investments in Artificial Intelligence (AI) and especially in (Deep) Machine Learning (ML). This led to breakthroughs in their practical ability to solve complex real-world problems impacting research and society at large. Instead, our ability to understand the fundamental mechanism behind these breakthroughs has slowed down because of their increased complexity, while in the past breakthroughs often emerged from foundational research. This questioned researchers about the necessity for a new theoretical framework able to help researchers catch up on this lag. One of the still not well understood mechanisms is the so-called over-parametrization, namely the ability of certain models to increase their generalization performance (reduce test error) when the number of parameters is above the interpolating threshold (zero training error). In this paper we will show that this phenomenon can be better understood using both known theories (surveying them in the process) and empirical evidences for both shallow and deep learning algorithms.

The Analysis of Enterprise Improvement in Global Commodity Price Prediction Based on Deep Learning

The article expects to solve the traditional econometric statistical model, shallow machine learning algorithm, and many limitations in learning the nonlinear relationship of related indicators affecting commodity futures price trend. This article proposes a neural network commodity futures price prediction model by the mixture of convolutional neural networks (CNN) and gated recurrent unit (GRU). Firstly, the dimension reduction algorithm of multidimensional data by principal component analysis (PCA) is used. Through linear transformation, the original variables with correlation are transformed into a set of new linear irrelevant variables, and the high-dimensional time series data of commodity futures are reduced. Secondly, the variable features are extracted from the CNN network module in the CNN-GRU model, and the GRU network module learns the periodicity and trend of the original data. Finally, the full connection layer outputs the forecast results of commodity futures price.

An exploratory study of underwater bolted connection looseness detection using percussion and a shallow machine learning algorithm

An exploratory study of underwater bolted connection looseness detection using percussion and a shallow machine learning algorithm

Fog-cloud based intrusion detection system using Recurrent Neural Networks and feature selection for IoT networks

Deep learning (DL) techniques are being widely researched for their effectiveness in detecting cyber intrusions against the Internet of Things (IoT). Time sensitive Critical Infrastructures (CIs) that rely on IoT require rapid detection of cyber intrusions close to the constrained devices in order to prevent service delays. Deep learning techniques perform better in detecting attacks compared to shallow machine learning algorithms and can be used for intrusion detection. However, communication overheads due to large volume of IoT data and computation requirements for deep learning models prevents effective application of deep learning models closer to the constrained devices. Existing IDS techniques are either based on shallow learning algorithms or not trained on relevant IoT datasets and furthermore not designed for distributed fog-cloud deployment. To counter these issues, we propose a novel fog-cloud based IoT intrusion detection framework which incorporates a distributed processing by splitting the dataset according to attack class and a feature selection step on time-series IoT data. This is followed by a deep learning Recurrent Neural Network (SimpleRNN and Bi-directional Long Short-Term Memory (LSTM)) for attack detection. The effectiveness of the proposed approach was evaluated using the high-dimensional BoT-IoT dataset which contains large volumes of realistic IoT attack traffic. Results show that feature selection methods significantly reduced the dataset size by 90% under the computation requirements without compromising on the attack detection ability. The models built on reduced dataset achieved higher recall rate compared to models trained on full feature set without loosing class differentiation ability. The SimpleRNN and Bi-LSTM models also did not suffer any underfitting or overfitting with the reduced feature space. The proposed deep learning based IoT intrusion detection framework is suitable for fog-cloud based deployment and can scale well even with large volumes of IoT data.