Artificial Intelligence XXXVI

In multi-label classification a datapoint can be labelled with more than one class at the same time. A common but trivial approach to multi-label classification is to train individual binary classifiers per label, but the performance can be improved by considering associations between the labels, and algorithms like classifier chains and RAKEL do this effectively. Like most machine learning algorithms, however, these approaches require accurate hyperparameter tuning, a computationally expensive optimisation problem. Tuning is important to train a good multi-label classifier model. There is a scarcity in the literature of effective multi-label classification approaches that do not require extensive hyperparameter tuning. This paper addresses this scarcity by proposing CascadeML, a multi-label classification approach based on cascade neural network that takes label associations into account and requires minimal hyperparameter tuning. The performance of the CasecadeML approach is evaluated using 10 multi-label datasets and compared with other leading multi-label classification algorithms. Results show that CascadeML performs comparatively with the leading approaches but without a need for hyperparameter tuning.

Support Vector Machines can achieve levels of accuracy comparable to those achieved by Artificial Neural Networks, but they are also slower to train. In this paper a new algorithm, called Purity Filtering, is presented, designed to filter training data for binary classification SVMs, in order to choose an approximation of the data subset that is more relevant to the training process.The proposed algorithm is parametrized so to allow a regulation of both spatial and temporal complexity, adapting to the needs and possibilities of each execution environment. A user-specified parameter, the purity, is used to indirectly regulate the number of filtered data, even though the algorithm has also been adapted to let the user directly specify the number of filtered data. Using this algorithm with real datasets, reductions up to 75% of training data (using only 25% of the data samples to train) were achieved with no major loss on the quality of classification.

Deep reinforcement learning has over the past few years shown great potential in learning near-optimal control in complex simulated environments with little visible information. Rainbow (Q-Learning) and PPO (Policy Optimisation) have shown outstanding performance in a variety of tasks, including Atari 2600, MuJoCo, and Roboschool test suite. Although these algorithms are fundamentally different, both suffer from high variance, low sample efficiency, and hyperparameter sensitivity that, in practice, make these algorithms a no-go for critical operations in the industry.On the other hand, model-based reinforcement learning focuses on learning the transition dynamics between states in an environment. If the environment dynamics are adequately learned, a model-based approach is perhaps the most sample efficient method for learning agents to act in an environment optimally. The traits of model-based reinforcement are ideal for real-world environments where sampling is slow and in mission-critical operations. In the warehouse industry, there is an increasing motivation to minimise time and to maximise production. In many of these environments, the literature suggests that the autonomous agents in these environments act suboptimally using handcrafted policies for a significant portion of the state-space.In this paper, we present The Dreaming Variational Autoencoder v2 (DVAE-2), a model-based reinforcement learning algorithm that increases sample efficiency, hence enable algorithms with low sample efficiency function better in real-world environments. We introduce the Deep Warehouse environment for industry-near testing of autonomous agents in logistic warehouses. We illustrate that the DVAE-2 algorithm improves the sample efficiency for the Deep Warehouse compared to model-free methods.

This paper explores how Learned Guidance Functions (LGFs)—a pre-training method used to smooth search landscapes—can be used as a fitness function for evolutionary algorithms. A new form of LGF is introduced, based on deep neural network learning, and it is shown how this can be used as a fitness function. This is applied to a test problem: unscrambling the Rubik’s Cube. Comparisons are made with a previous LGF approach based on random forests, and with a baseline approach based on traditional error-based fitness.

Longitudinal datasets contain repeated measurements of the same variables at different points in time, which can be used by researchers to discover useful knowledge based on the changes of the data over time. Monotonic relations often occur in real-world data and need to be preserved in data mining models in order for the models to be acceptable by users. We propose a new methodology for detecting monotonic relations in longitudinal datasets and applying them in longitudinal classification model construction. Two different approaches were used to detect monotonic relations and include them into the classification task. The proposed approaches are evaluated using data from the English Longitudinal Study of Ageing (ELSA) with 10 different age-related diseases used as class variables to be predicted. A gradient boosting algorithm (XGBoost) is used for constructing classification models in two scenarios: enforcing and not enforcing the constraints. The results show that enforcement of monotonicity constraints can consistently improve the predictive accuracy of the constructed models. The produced models are fully monotonic according to the monotonicity constraints, which can have a positive impact on model acceptance in real world applications.

Whereas most work in reinforcement learning (RL) ignores the structure or relationships between actions, in this paper we show that exploiting structure in the action space can improve sample efficiency during exploration. To show this we focus on concurrent action spaces where the RL agent selects multiple actions per timestep. Concurrent action spaces are challenging to learn in especially if the number of actions is large as this can lead to a combinatorial explosion of the action space.This paper proposes two methods: a first approach uses implicit structure to perform high-level action elimination using task-invariant actions; a second approach looks for more explicit structure in the form of action clusters. Both methods are context-free, focusing only on an analysis of the action space and show a significant improvement in policy convergence times.

A successful dialogue requires that the participants have a shared understanding of what they are trying to achieve, individually and collectively. This coordination can be achieved if both recognise the type of dialogue in which they are engaged. We focus on two particular dialogue types, action persuasion and deliberation dialogues, which are often conflated because they share similar speech acts. Previously, a clear distinction was made between the two in terms of the different pre- and post-conditions used for the speech acts within these dialogues. This prior work gave formal specifications of the dialogue moves within the dialogues but offered no evaluation through implementation. In this paper, we present an implementation to demonstrate that the two dialogue types described in this way can be realised in software to support focussed communication between autonomous agents. We provide the design and implementation details of our new tool along with an evaluation of the software. The tool we have produced captures the distinctive features of each of the two dialogue types, to make plain their differences and to validate the speech acts for use in practical scenarios.

Patients with Learning Disabilities (LD) have substantial and unmet healthcare needs, and previous studies have highlighted that they face both health inequalities and worse outcomes than the general population. Primary care practitioners are often the first port-of-call for medical consultations, and one issue faced by LD patients in this context is the very limited time available during consultations - typically less than ten minutes. In order to alleviate this issue, we propose a digital communication aid in the form of an ontology-based medical questionnaire that can adapt to a patient’s medical context as well as their accessibility needs (physical and cognitive). The application is intended to be used in advance of a consultation so that a primary care practitioner may have prior access to their LD patients’ self-reported symptoms. This work builds upon and extends previous research carried out in the development of adaptive medical questionnaires to include interactive and interface functionalities designed specifically to cater for patients with potentially complex accessibility needs. A patient’s current health status and accessibility profile (relating to their impairments) is used to dynamically adjust the structure and content of the medical questionnaire. As such, the system is able to significantly limit and focus questions to immediately relevant concerns while discarding irrelevant questions. We propose that our ontology-based design not only improves the relevance and accessibility of medical questionnaires for patients with LDs, but also provides important benefits in terms of medical knowledge-base modularity, as well as for software extension and maintenance.

With the amount of information that exists online, it is impossible for a student to find relevant information or stay focused on the domain under study. Research showed that search engines have deficiencies that might prevent students from finding relevant information. To this end, this research proposes a technical solution that takes the personal search history of a student into consideration and provides a holistic view of the domain under study. Based on algorithmic approaches to assert semantic similarity, the proposed framework makes use of a user interface to dynamically assist students through aggregated results and wordcloud visualizations. The effectiveness of our approach is finally evaluated through the use of commonly used datasets and compared in line with existing research.

Automated planning techniques are increasingly exploited in real-world applications, thanks to their flexibility and robustness. Hybrid domains, those that require to reason both with discrete and continuous aspects, are particularly challenging to handle with existing planning approaches due to their complex dynamics. In this paper we present a general approach that allows to combine the strengths of automated planning and control systems to support reasoning in hybrid domains. In particular, we propose an architecture to integrate Model Predictive Control (MPC) techniques from the field of control systems into an automated planner, to guide the effective exploration of the search space.

The recently introduced Tsetlin Machine (TM) has provided competitive pattern recognition accuracy in several benchmarks, however, requires a 3-dimensional hyperparameter search. In this paper, we introduce the Multigranular Tsetlin Machine (MTM). The MTM eliminates the specificity hyperparameter, used by the TM to control the granularity of the conjunctive clauses that it produces for recognizing patterns. Instead of using a fixed global specificity, we encode varying specificity as part of the clauses, rendering the clauses multigranular. This makes it easier to configure the TM because the dimensionality of the hyperparameter search space is reduced to only two dimensions. Indeed, it turns out that there is significantly less hyper-parameter tuning involved in applying the MTM to new problems. Further, we demonstrate empirically that the MTM provides similar performance to what is achieved with a finely specificity-optimized TM, by comparing their performance on both synthetic and real-world datasets.

Workflows are core part of every modern organization ensuring smooth running operations, task consistency and process automation. Dynamic workflows are being used increasingly due to their flexibility in a working environment where they minimize mundane tasks like long-term maintenance and increase productivity by automatically responding to changes and introducing new processes. Constant changes within unstable environments where information may be sparse, inconsistent and uncertain can create a bottleneck to a workflow in predicting behaviours effectively. Within a business environment, automatic applications like customer support, complex incidents can be regarded as instances of a dynamic process since mitigation policies have to be responsive and adequate to any case no matter its unique nature. Support engineers work with any means at their disposal to solve any emerging case and define a custom prioritization strategy, to achieve the best possible result. This paper describes a novel workflow architecture for heavy knowledge-related application workflows to address the tasks of high solution accuracy and shorter prediction resolution time. We describe how policies can be generated against cases deriving from heterogeneous workflows to assist experts and domain-specific reusable cases can be generated for similar problems. Our work is evaluated using data from real business process workflows across a large number of different cases and working environments.

Violence detection is becoming increasingly relevant in many areas such as for automatic content filtering, video surveillance and law enforcement. Existing datasets and methods discriminate between violent and non-violent scenes based on very abstract definitions of violence. Available datasets, such as “Hockey Fight” and “Movies”, only contain fight versus non-fight videos; no weapons are discriminated in them. In this paper, we focus explicitly on weapon-based fighting sequences and propose a new dataset based on the popular action-adventure video game Grand Theft Auto-V (GTA-V). This new dataset is called “Weapon Violence Dataset” (WVD). The choice for a virtual dataset follows a trend which allows creating and labelling as sophisticated and large volume, yet realistic, datasets as possible. Furthermore, WVD also avoids the drawbacks of access to real data and potential implications. To the best of our knowledge no similar dataset, that captures weapon-based violence, exists. The paper evaluates the proposed dataset by utilising local feature descriptors using an SVM classifier. The extracted features are aggregated using the Bag of Visual Word (BoVW) technique to classify weapon-based violence videos. Our results indicate that SURF achieves the best performance.

In the service industry, having an efficient resource plan is of utmost importance for operational efficiency. An accurate forecast of demand is crucial in obtaining a resource plan which is efficient. In this paper, we present a real world application of an AI forecasting model for vehicle volumes forecasting in service stations. We improve on a previously proposed approach by intelligently tuning the hyper parameters of the prediction model, taking into account the variability of the vehicle volume data in a service station. In particular, we build a Genetic algorithm based model to find the topology of the neural network and also to tune additional parameters of the prediction model that is related to data filtration, correction and feature selection. We compare our results with the results from ad hoc parameter settings of the model from previous work and show that the combined genetic algorithm and neural network based approach further improves forecasting accuracy which helps service stations better manage their resource requirements.

Physical activity (PA) is a key component in the treatment of a range of chronic health conditions. It is therefore important for researchers and clinicians to accurately assess and monitor PA. Although advances in wearable technology have improved this, there is a need to investigate PA in greater depth than the sum of its total parts. Specifically, linking deep PA data to patient outcomes offers a valuable, and unexplored use for wearable devices. As a result, this paper extracts useful features from accelerometer data (Actigraph GT3X Link), and applies machine learning algorithms to predict daily pain and stiffness. This was applied to a population of 30 arthritis patients and 15 healthy volunteers. Participants were provided with an Actigraph and asked to wear it continuously for 28 days. Results demonstrate that it is possible to predict both pain and stiffness of patients using the extracted accelerometer features.

A mechanism is presented for classifying phonocardiograms (PCGs) by interpreting PCGs as time series and using the concept of motifs, times series subsequences that are good discriminators of class, to support nearest neighbour classification. A particular challenge addressed by the work is that PCG time series are large which renders exact motif discovery to be computationally expensive; it is not realistic to compare every candidate time series subsequence with every other time series subsequence in order to discover exact motifs. Instead, a mechanism is proposed the firstly makes use of the cyclic nature of PCGs and secondly adopts a novel time series pruning mechanism. The evaluation, conducted using a canine PCG dataset, illustrated that the proposed approach produced the same classification accuracy but in a significantly more efficient manner.

The Animal and Plant Health Agency (APHA) conducts post-mortem examinations (PMEs) of farm animal species as part of routine scanning surveillance for new and re-emerging diseases that may pose a threat to animal and public health. This paper investigates whether relevant veterinary medical terms can be automatically identified in the free-text summaries entered by Veterinary Investigation Officers (VIOs) on the PME reports. Two natural language processing tasks were performed: (1) named entity recognition, where terms within the free-text were mapped to concepts in the Unified Medical Language System (UMLS) Metathesaurus; and (2) semantic similarity and relatedness also using UMLS. For this pilot study, we focused on two diagnostic codes: salmonellosis (S. Dublin) and Pneumonia NOS (Not Otherwise Specified). The outputs were manually evaluated by VIOs. The results highlight the potential value of natural language processing to identify key concepts and pertinent veterinary medical terms that can be used for scanning surveillance purposes using large, free-text data. We also discuss issues resulting from the inherent bias of UMLS to human medical terms and its use in animal health monitoring.

Creativity is often defined as the creation of something novel through the use of imagination. But for all artists, creativity is also the exploration of new and unknown areas within their specific art. Is it possible to stimulate creativity through a system that creates inspiring original music, and that is also able to learn the personal tastes of its user? Within the project, an evolutionary approach was used in an attempt to stimulate musical creativity by supplying a composer with software that can compose short musical patterns called riffs. The software, called GenMuse, evolves populations of riffs, and makes use of a feed-forward artificial neural network to learn how to autonomously evaluate the evolved riffs to satisfy the tastes of the composer. The results show that the approach is worthy of further investigation. The genetic algorithm produced interesting results that, according to our evaluation parameters, could be included “as is” in a musical composition, and the neural network was able to evaluate the riffs with a good success ratio.

This project involved the use of an interactive Genetic Algorithm (iGA) with an electroencephalogram (EEG)-based fitness function to create paintings in the style of Piet Mondrian, a Dutch painter who used geometric elements in his later paintings. Primary data for the prototype was gathered by analysis of twenty-seven existing Mondrian paintings. An EEG gaming headset was used to read EEG signals, which were transmitted by Bluetooth to an Arduino running an iGA. These values were used as the iGA fitness function. The data was sent to a PC running Processing to display the artwork. The resultant displayed artwork evolves to favour higher attention and meditation levels, which are considered to represent greater mindfulness. The process ends when the observer identifies a piece of art they would like to keep. However, convergence of the algorithm is difficult to test as many parameters can affect the process. A number of issues arising from the research are discussed and further work is proposed.

The increasing data traffic inside buildings requires maintaining good cellular network coverage for indoor mobile users. Passive In-building Distributed Antenna System (IB-DAS) is one of the most efficient methods to provide an indoor solution that meets the signal strength requirements. It is a network of spatially distributed antennas in a building connected to telephone rooms which are then connected to a Base Transmission Station (BTS). These connections are established through passive coaxial cables and splitters. The design of IB-DAS is considered to be challenging due to the power-sharing property resulting in two contradicting objectives: minimizing the power usage at the BTS (long-term cost) and minimizing the design components cost (short-term cost). Different attempts have been made in the literature to solve this problem. Some of them are either lacking the consideration of all necessary aspects or facing scalability issues. Additionally, most of these attempts translate the IB-DAS design into a mono-objective problem, which leads to a challenging task of determining a correct combined objective function with justified weighting factors associated with each objective. Moreover, these approaches do not produce multiple design choices which may not be satisfactory in practical scenarios. In this paper, we propose a multi-objective algorithm for designing IB-DAS. The experimental results show the success of this algorithm to achieve our industrial partner’s requirement of providing different design options that cannot be achieved using mono-objective approaches.

The aim of this paper is to investigate different machine learning based forecasting techniques for forecasting of blood pressure and heart rate. Forecasting of blood pressure could potentially help a clinician to take preventative steps even before dangerous medical situations occur. This paper examines forecasting blood pressure 30 min in advance. Univariate and multivariate forecast models are considered. Different forecast strategies are also considered. To compare different forecast strategies, LSTM and BI-LSTM machine learning algorithms were included. Then univariate and multivariate LSTM, BI-LSTM and CNN machine learning algorithms were compared using the two best forecasting strategies. Comparative analysis between forecasting strategies suggest that MIMO and DIRMO forecast strategies provide the best accuracy in forecasting physiological time series data. Results also appear to show that multivariate forecast models for blood pressure and heart rate are more reliable compared to blood pressure alone. Comparative analysis between MIMO and DIRMO forecasting strategies appear to show that DIRMO is more reliable for both univariate and multivariate cases. Results also appear to show that the forecast model that uses BI-LSTM with the DIRMO strategy is the best overall.

Forecasting of financial time series is challenging due to its non-linear and non-stationary characteristics. Due to limitations of traditional time series models, it is difficult to forecast financial time series such as stock price and stock index. Hence, we used ensemble of time series decomposition-based models (such as Discrete Wavelet Transform, Empirical Mode Decomposition and Variational Mode Decomposition) and machine learning models (such as Artificial Neural Network and Support Vector Regression) for forecasting the close price of 25 major stock indices for a period of 10 years ranging from January 1, 2009 to December 31, 2018. Decomposition models are used to disaggregate the time series into various subseries and machine learning models are used for forecasting each subseries. The forecasted subseries are then aggregated to obtain the final forecast. The performance of the models was evaluated using Root Mean Square Error and was validated statistically using Wilcoxon Signed Rank Test. We found that the performance of ensemble models better than traditional machine learning models.

k Nearest Neighbour classification techniques, where $$k=1$$, coupled with Dynamic Time Warping (DTW) are the most effective and most frequently used approaches for time series classification. However, because of the quadratic complexity of DTW, research efforts have been directed at methods and techniques to make the DTW process more efficient. This paper presents a new approach to efficient DTW, the Sub-Sequence-Based DTW approach. Two variations are considered, fixed length sub-sequence segmentation and fixed number sub-sequence segmentation. The reported experiments indicate that the technique improvs efficiency, compared to standard DTW, without adversely affecting effectiveness.

Organisations face growing legal requirements and ethical responsibilities to ensure that decisions made by their intelligent systems are explainable. However, provisioning of an explanation is often application dependent, causing an extended design phase and delayed deployment. In this paper we present an explainability framework formed of a catalogue of explanation methods, allowing integration to a range of projects within a telecommunications organisation. These methods are split into low-level explanations, high-level explanations and co-created explanations. We motivate and evaluate this framework using the specific case-study of explaining the conclusions of field engineering experts to non-technical planning staff. Feedback from an iterative co-creation process and a qualitative evaluation is indicative that this is a valuable development tool for use in future company projects.

The prime concern for a business organization is to supply quality services to the customers without any delay or interruption so to establish a good reputation among the customer’s and competitors. On-time delivery of a customers order not only builds trust in the business organization but is also cost effective. Therefore, there is a need is to monitor complex business processes though automated systems which should be capable during execution to predict delay in processes so as to provide a better customer experience. This online problem has led us to develop an automated solution using machine learning algorithms so as to predict possible delay in business processes. The core characteristic of the proposed system is the extraction of generic process event log, graphical and sequence features, using the log generated by the process as it executes up to a given point in time where a prediction need to be made (referred to here as cut-off time); in an executing process this would generally be current time. These generic features are then used with Support Vector Machines, Logistic Regression, Naive Bayes and Decision trees to predict the data into on-time or delayed processes. The experimental results are presented based on real business processes evaluated using various metric performance measures such as accuracy, precision, sensitivity, specificity, F-measure and AUC for prediction as to whether the order will complete on-time when it has already been executing for a given period.

This paper presents an investigation of speech emotion systems and how the accuracy can be further improved by exploring machine learning algorithms and hybrid solutions. The accuracy of machine learning algorithms and speech noise reduction techniques are investigated on an embedded system. Research suggests improvements could be made to the feature selection from speech signals and pattern recognition algorithms for emotion recognition. The system deployed to perform the experiments is EmotionPi, using the Raspberry Pi 3 B+. Pattern recognition is investigated by using K-Nearest Neighbour (K-NN), Support Vector Machine (SVM), Random Forest Classifier (RFC), Multi-Layer Perception (MLP) and Convolutional Neural Networks (CNN) algorithms. Experiments are conducted to determine the accuracy of the speech emotion system using the speech database and our own recorded dataset. We propose a hybrid solution which has proven to increase the accuracy of the emotion recognition results. Results obtained from testing, show the system needs to be trained using real cases rather than using speech databases (as it is more accurate in detecting the user’s emotion).

Automated defect prediction of large and complex software systems is a challenging task. However, by utilising correlated quality metrics, a defect prediction model can be devised to automatically predict the defects in a software system. The robustness and accuracy of a prediction model is highly dependent on the selection of contributing and non-contributing features. Hence, in this regard, the contribution of this paper is twofold, first it separates those features which are contributing towards the development of a defect in a software component from those which are non-contributing features. Secondly, a logistic regression and Ensemble Bagged Trees-based prediction model are applied on the contributing features for accurately predicting a defect in a software component. The proposed models are compared with the most recent scheme in the literature in terms of accuracy and area under the curve (AUC). It is evident from the results and analysis that the performance of the proposed prediction models outperforms the schemes in the literature.

Inter-turn short circuit of the stator is one of the most common faults of an induction motor that degrades its performance and ultimately causes it to break down. To avoid unexpected breakdown, causing an industrial process to halt, it is desirable to continuously monitor the motor’s operation using an automated system that can differentiate normal from faulty operation. However, such automated systems usually require large datasets containing enough examples of normal and faulty characteristics of the motor to be able to detect abnormal behavior. The aim of this paper is to present some ways to extract such information or features from the available sensor signals data like motor currents, voltages and vibration, to enable a machine learning based fault detection system to discern normal operation from faulty operation with minimal training data.

Identifying types of file fragments in isolation from their context is an essential task in digital forensic analysis and can be done with several methods. One common approach is to extract various types of features from file fragments as inputs for classification algorithms. However, this approach suffers from dimensionality curse as the number of the extracted features is too high, which causes the learning and classification to be both inefficient and inaccurate. This paper proposes a hybrid method to address this issue by using filters and wrappers to significantly reduce the number of features and also improve the accuracy of file type classification. First, it uses and combines three appropriate filters to filter out a large number of irrelevant and/or less important features, and then some wrappers to reduce the number of features further to the most salient ones. Our method was tested on some benchmark datasets - GovDocs, and the experimental results indicated that our method was able to not only reduce the number of features from 66,313 to 11–32, but also improve the accuracy of the classification, compared with other methods that used all the features.

The demand for photovoltaic cells has been increasing exponentially in the past few years because of its potential for generating clean electricity. Yet, due to low efficiency, this technology has not demonstrated complete reliability and poses tremendous amount of constraints even after the possibility of substantial power outputs. The concept of multi-junction solar cell has provided partial solution to this problem. Since the multi-junction solar cell was developed, its optimization has posed a great challenge for the entire community. The present study has been conducted on Si tandem cell, which is a two-junction three-layered solar cell. Silicon (Si) tandem cell was one of the initial developments in the domain of multi-junction solar cells and is most commercially fabricated photovoltaic cell. In this paper, the optimization challenge of multi-junction solar cells has been attempted with the use of Artificial Neural Network (ANN) technique. Artificial Neural Network was trained by using Bayesian Regularization algorithm, and used. Input parameters were taken as spectral power density, temperature and thickness of the layers of cells. Voltage of the cell was kept as a biasing input, and the output parameter was taken to be current density. I-V characteristics were plotted which was further used to calculate the open-circuit voltage (Voc), Fill Factor of the cell (FF), short circuit current density (Jsc) and Maximum Power Point (MPP). The output generated by the trained model of ANN has been compared with the values generated by more than a million iteration of the solar cell model. The implementation of this algorithm on any model of the multi-junction solar cell can lead to the development of highly efficient solar cells. Thus, with due consideration of physical constraints of the environment where it is to be installed; maximum amount efficiency can be achieved.

Intrusion detection is the ability to mitigate attacks and block new threats. In this paper, we deal with intrusion detection as a pattern classification problem where a connection is defined as a set of attributes. The latter forms a pattern that should be assigned to one of existing classes. The problem is to identify the given connection as a normal event or attack. We propose a stochastic local search method for feature selection where the aim is to select the set of significant attributes to be used in the classification task. The proposed approach is validated on the well-known NLS-KDD dataset and compared with some existing techniques. The results are interesting and show the efficiency of the proposed approach for intrusion detection.

Artificial Intelligence (AI) has been around for many years and plays a vital role in developing automatic systems that require decision using a data- or model-driven approach. Smart homes are one such system; in them, AI is used to recognize user activities, which is a fundamental task in smart home system design. There are many approaches to this challenge, but data-driven activity recognition approaches are currently perceived the most promising to address the sensor selection uncertainty problem. However, a smart home using a data-driven approach exclusively cannot immediately provide its new occupant with the expected functionality, which has reduced the popularity of the data-driven approach. This paper proposes an approach to develop an integrated personalized system using a user-centric approach comprising survey, simulation, activity recognition and transfer learning. This system will optimize the behaviour of the house using information from the user’s experience and provide required services. The proposed approach has been implemented in a smart home and validated with actual users. The validation results indicate that users benefited from smart features as soon as they move into the new home.

The automatic reconstruction of the patient’s treatment lines from their Electronic Health Records (EHRs) is a significant step towards improving the quality and the safety of the healthcare deliveries. With the recent rapid increase in the adaption of EHRs and the rapid development of computational science, we can discover new insights from the information stored in EHRs. However, this is still a challenging task, being unstructured data analysis one of them. In this paper, we focus on the most common challenges for reconstructing the patient’s treatment lines, which are the Named Entity Recognition (NER), temporal relation identification and the integration of structured results. We introduce our Natural Language Processing (NLP) framework, which deals with the aforementioned challenges. In addition, we focus on a real use case of patients, suffering from lung cancer to extract patterns associated with the treatment of the disease that can help clinicians to analyze toxicities and patterns depending on the lines of treatments given to the patient.

In Voxel-Based Morphometry (VBM), spatial normalisation is a major process which transforms images into a standard space and is often referred to as co-registration. This project is a comparison and observation of differences in the performance, measured as the overlap between images, of two co-registration algorithms used in VBM on human brain Magnetic Resonance Imaging (MRI) data. Here we show differences between genders and algorithms on specific regions of the brain using grey matter segments and unsegmented MRI images. Results show that there are significant differences in the overlap of regions depending on the algorithm which may be considered in addition to current knowledge on the subject. Importantly, we are interested in investigating what these differences mean to published and on-going research as well as observing whether said difference spans all the way to the Parahippocampal Gyrus and other important regions associated with psychological related diseases.

The paper presents an algorithm providing creativity support to journalists. It suggests analogical transfer of news angles from reports written about different events than the one the journalist is working on. The problem is formulated as a matching problem, where news reports with similar wordings from two events are matched, and unmatched reports from previous cases are selected as candidates for a news angle transfer. The approach is based on document similarity measures for matching and selection of transferable candidates. The algorithm has been tested on a small data set and show that the concept may be viable, but needs more exploration and evaluation in journalistic practice.

This paper explores the use of an interactive Genetic Algorithm for creating a piece of visual art intended to assist in promoting the state of mindfulness. This is determined by a Bluetooth gaming electroencephalography (EEG) headset as the fitness function. The visual display consisted of an infinity mirror with over two hundred Neopixels with fade times and colour of zones controlled by two Arduinos running the software. Whilst we have observed some convergence of solutions, the results and user observations raised some interesting questions about how this strategy might be improved.

With an ever growing demand for providing AI solutions within business there is a tendency to expect end to end standardised solutions to problems. These solutions are expected to be accurate and to be seamlessly integrated within existing business processes. However, achieving higher accuracy could be detrimental not only to explainability (if a blackbox solution is provided) but also need to be accepted and used within existing business processes. This paper describes a Case-Based Reasoning (CBR) solution to a real problem within a telecommunications company together with the reasoning behind selecting this particular approach. The solution has been integrated within existing business processes which has been a real challenge besides satisfying all the technical ability criteria.

Emotion detection is an important step for recognizing a person’s mental state. A physiological signal, Electroencephalogram (EEG) is analyzed to detect human emotion with promising results. The cost of information gathering and lack of number of participants incur a limitation on the size of EEG data set. The deficiency in acquired EEG data set makes it difficult to estimate mental states with deep learning models as it requires a larger size of the training data set. In this paper, we propose a novel data augmentation method to address challenges due to scarcity of EEG data for training deep learning models such as Recurrent Neural Network - Long Short Term Memory (RNN-LSTM). To find the performance of mental state estimator such models are applied before and after proposed data augmentation. Experimental results demonstrate that data augmentation improves the performance of mental state estimator with an accuracy of 98%.

We aim to classify financial discrepancies between actual and forecasted performance into categories of commentaries that an analyst would write when describing the variation. We propose analyzing time series in order to perform the classification. Two time series classification algorithms – 1-nearest neighbour with dynamic time warping (1-NN DTW) and time series forest – and long short-term memory (LSTM) networks are compared to common machine learning algorithms. We investigate including supporting datasets such as customer sales data and inventory. We apply data augmentation with noise as an alternative to random oversampling. We find that LSTM and 1-NN DTW provide the best results. Including sales data has no effect but inventory data improves the predictive power of all models examined. Data augmentation has a slight improvement for some models over random oversampling.

In precision agriculture (PA), soil sampling and testing operation is prior to planting any new crop. It is an expensive operation since there are many soil characteristics to take into account. This paper gives an overview of soil characteristics and their relationships with crop yield and soil profiling. We propose an approach for predicting soil pH based on nearest neighbour fields. It implements spatial radius queries and various regression techniques in data mining. We use soil dataset containing about 4, 000 fields profiles to evaluate them and analyse their robustness. A comparative study indicates that LR, SVR, and GBRT techniques achieved high accuracy, with the $$R_2$$ values of about 0.718 and $$MAE$$ values of 0.29. The experimental results showed that the proposed approach is very promising and can contribute significantly to PA.

Policy making involves an extensive research phase during which existing policies which are similar to the one under development need to be retrieved and analysed. This phase is time-consuming for the following reasons: (i) there is no unified format for policy documents; (ii) there is no unified repository of policies; and (iii) there is no retrieval system designed for querying any repositories which may exist. This creates an information overload problem for policy makers who need to be aware of other policy documents in order to inform their own. The goal of this work is to introduce a novel application area for studying information retrieval models: the information seeking phase of policy design, applied to life-long learning policy-making. In this paper, we address this problem by developing a common representation for policy documents, informed by domain experts, in order to facilitate their indexing and retrieval by users. This position paper highlights the research questions that we aim to answer in our future work and the dataset that we intend to use to do so. Our main contribution is the creation of a unified dataset of policy interventions which can be used for highly specialised information retrieval tasks, and will be released in order to provide the field with the first unified repository of policy interventions in adult education.

The current advances of computational power and storage allow more models to be created and stored from significant data resources. This progress opens the opportunity to re-cycle and re-use such models in similar exercises. The evaluation of the machine learning algorithms and selection of an appropriate classifier from an existing collection of classifiers are still challenging tasks. In most cases, the decision of selecting the classifier is left to the user. When the selection is not performed accurately, the outcomes can have unexpected performance results. Classification algorithms aim to optimise some of the distinct objectives such as minimising misclassification error, maximising the accuracy, or maximising the model quality. The right choice for each of these objectives is critical to the quality of the classifier selected. This work aims to study the use of a multi-objective method that can be undertaken to find a set of suitable classifiers for a problem at hand. In this study, we applied seven classifiers on mental health data sets for classifier selection in terms of correctness and reliability. The experimental results suggest that this approach is useful in finding the best trade-off among the objectives of selecting a suitable classifier framework.