Abstract Time-series forecasts underpin decision-making processes in a wide range of application domains. Recently it has been shown that these processes can be strengthened by conformal prediction, a framework that allows adding prediction intervals to point forecasts. The prediction intervals quantify the uncertainty of a predictive model with mathematical coverage guarantees, giving the user a range of scenarios to consider. However, applying conformal prediction to time-series tasks is not trivial. This is either because the exchangeability condition the framework places on the data is violated, or because the framework only allows for one-step-ahead univariate forecasts. In this article we combine two existing methods derived from conformal prediction, one built for multi-target regression and one designed to handle non-exchangeable data. The resulting method, called non-exchangeable multi-target conformal prediction (nmtCP) produces provably robust prediction regions for multi-step ahead multidimensional time-series forecasts, meaning that the miscoverage rate is bound. Additionally, nmtCP is computationally efficient and easy to implement. Due to its model-agnostic nature, nmtCP can be used on top of any time-series model that produces point forecasts. A theoretical analysis proves the method’s robustness while experiments on real-world data sets give insights into its practical behavior and performance.

Solar energy is one of the most promising renewable energy sources. However, it is highly variable, which motivates the development of accurate methods for forecasting the generated solar power, to facilitate its integration into the power grid. In this paper, we consider the task of predicting the half-hourly PV solar power for the next day, from three data sources: previous solar power, previous weather data and weather forecast for future days. We investigate the potential of LSTM and CNN neural networks, and also propose the new method LSTM-Conv, which leverages the strengths of LSTM to learn temporal dependencies and of CNN to learn useful features. The evaluation is conducted on two solar power datasets, for two years. The results showed that LSTM-Conv was the most accurate method, outperforming LSTM, CNN, MLP, RNN and a persistence baseline. The best result on both datasets was achieved by LSTM-Conv with weather data. The use of weather data considerably improved the results, highlighting its importance for day-ahead solar power forecasting.

Obstructive sleep apnea (OSA) is a sleep disorder that causes partial or complete cessation of breathing during an individual's sleep. Various methods have been proposed to automatically detect OSA events, but little work has focused on predicting such events in advance, which is useful for the development of devices that regulate breathing during a patient's sleep. We propose four methods for sleep apnea prediction based on convolutional and long short-term memory neural networks (1D-CNN, ConvLSTM, 1D-CNN-LSTM and 2D-CNN-LSTM), which use raw data from three respiratory signals (nasal flow, abdominal and thoracic) sampled at 32 Hz, without any human-engineered features. We predict OSA (apnea or hypopnea) and normal breathing events 30 seconds ahead using the prior 90 seconds' data. Our results on a dataset containing over 46,000 examples from 1,507 subjects show that all four models achieved promising accuracy ( 81%). The 1D-CNN-LSTM and 2D-CNN-LSTM were the best two performing models with accuracy, sensitivity and specificity over 83%, 81% and 85% respectively. These results show that OSA events can be accurately predicted in advance based on respiratory signals, opening up opportunities for the development of devices to preemptively regulate the airflow to sleepers to avoid these events. Furthermore, we demonstrate good prediction performance even when respiratory signals are downsampled by a factor of 32, to 1 Hz, for which our proposed 1D-CNN-LSTM achieved 82.94% accuracy, 81.25% sensitivity and 84.63% specificity. This robustness to low sampling frequencies allows our algorithms to be implemented in devices with low storage capacity, making them suitable for at-home environments.

Online programming courses can provide detailed automatic feedback for code that fails to meet various test conditions, but novice students often struggle with syntax errors and are unable to write valid testable code. Even for very simple exercises, the range of incorrect code can be surprising to educators with mastery of a programming language. This research paper presents an analysis of the error messages from code run by students in an introductory Python~3 programming course, participated in by 8680 primary and high-school students from 680 institutions. The invalid programs demonstrate a wide diversity of mistakes: even for a one-line "Hello World!'' exercise there were 115 unique invalid programs. The most common errors are identified and compared to the topics introduced in the course. The most generic errors in selected exercises are investigated in greater detail to understand the underlying causes. While the majority of students attempting an exercise reach a successful outcome, many students encounter at least one error in their code. Of these, many such errors indicate basic mistakes, such as unquoted string literals, even in exercises late in the course for which some proficiency of earlier concepts is assumed. These observations suggest there is significant scope to provide greater reinforcement of students' understanding of earlier concepts.

Abstract Machine learning on trees has been mostly focused on trees as input. Much less research has investigated trees as output, which has many applications, such as molecule optimization for drug discovery, or hint generation for intelligent tutoring systems. In this work, we propose a novel autoencoder approach, called recursive tree grammar autoencoder (RTG-AE), which encodes trees via a bottom-up parser and decodes trees via a tree grammar, both learned via recursive neural networks that minimize the variational autoencoder loss. The resulting encoder and decoder can then be utilized in subsequent tasks, such as optimization and time series prediction. RTG-AEs are the first model to combine three features: recursive processing, grammatical knowledge, and deep learning. Our key message is that this unique combination of all three features outperforms models which combine any two of the three. Experimentally, we show that RTG-AE improves the autoencoding error, training time, and optimization score on synthetic as well as real datasets compared to four baselines. We further prove that RTG-AEs parse and generate trees in linear time and are expressive enough to handle all regular tree grammars.

In this paper, we present SSDNet, a novel deep learning approach for time series forecasting. SSDNet combines the Transformer architecture with state space models to provide probabilistic and interpretable forecasts, including trend and seasonality components and previous time steps important for the prediction. The Transformer architecture is used to learn the temporal patterns and estimate the parameters of the state space model directly and efficiently, without the need for Kalman filters. We comprehensively evaluate the performance of SSDNet on five data sets, showing that SSDNet is an effective method in terms of accuracy and speed, outperforming state-of-the-art deep learning and statistical methods, and able to provide meaningful trend and seasonality components.

Temporal Convolutional Neural Networks (TCNNs) have been applied for various sequence modelling tasks including time series forecasting. However, TCNNs may require many convolutional layers if the input sequence is long and are not able to provide interpretable results. In this paper, we present TCAN, a novel deep learning approach that employs attention mechanism with temporal convolutions for probabilistic forecasting, and demonstrate its performance in a case study for solar power forecasting. TCAN uses the hierarchical convolutional structure of TCNN to extract temporal dependencies and then uses sparse attention to focus on the important timesteps. The sparse attention layer of TCAN enables an extended receptive field without requiring a deeper architecture and allows for interpretability of the forecasting results. An evaluation using three large solar power data sets demonstrates that TCAN outperforms several state-of-the-art deep learning forecasting models including TCNN in terms of accuracy. TCAN requires less number of convolutional layers than TCNN for an extended receptive field, is faster to train and is able to visualize the most important timesteps for the prediction.

MOTIVATION: Many 'automated gating' algorithms now exist to cluster cytometry and single cell sequencing data into discrete populations. Comparative algorithm evaluations on benchmark datasets rely either on a single performance metric, or a few metrics considered independently of one another. However, single metrics emphasise different aspects of clustering performance and do not rank clustering solutions in the same order. This underlies the lack of consensus between comparative studies regarding optimal clustering algorithms and undermines the translatability of results onto other non-benchmark datasets. RESULTS: We propose the Pareto fronts framework as an integrative evaluation protocol, wherein individual metrics are instead leveraged as complementary perspectives. Judged superior are algorithms that provide the best trade-off between the multiple metrics considered simultaneously. This yields a more comprehensive and complete view of clustering performance. Moreover, by broadly and systematically sampling algorithm parameter values using the Latin Hypercube sampling method, our evaluation protocol minimises (un)fortunate parameter value selections as confounding factors. Furthermore, it reveals how meticulously each algorithm must be tuned in order to obtain good results, vital knowledge for users with novel data. We exemplify the protocol by conducting a comparative study between three clustering algorithms (ChronoClust, FlowSOM and Phenograph) using four common performance metrics applied across four cytometry benchmark datasets. To our knowledge, this is the first time Pareto fronts have been used to evaluate the performance of clustering algorithms in any application domain. AVAILABILITY: Implementation of our Pareto front methodology and all scripts to reproduce this article are available at https://github.com/ghar1821/ParetoBench.

We consider the task of predicting the solar power generated by a photovoltaic system, one-step ahead, from previous half-hourly photovoltaic power data. We propose a range of strategies for constructing static and dynamic heterogeneous ensembles and conduct an extensive evaluation using data for two years from two Australian solar power plants. We analyse the performance of the proposed static and dynamic ensembles and compare them with classical ensemble methods (bagging, boosting and random forest) and a baseline, showing an improved performance. The best result was achieved by the dynamic ensembles DEPast and DEPast+Future, in conjunction with the Peer Check algorithm for selecting base learners for inclusion in the dynamic ensembles.

Obstructive sleep apnea is a breathing disorder affecting 2-4% of the adult population. It is characterized by periods of reduced breathing (hypopnea) or no breathing (apnea). Several machine learning algorithms have been proposed to automatically classify sleep apnea events, but little work has been done on predicting such events in advance, which is important for the treatment of sleep apnea, and especially for the development of auto-adjusting airway pressure devices to maintain continuous airflow during sleep. In this paper, we propose three methods for predicting sleep apnea events, based on convolution neural networks and Markov chains. Specifically, we use data from respiratory signals (nasal flow, abdominal and thoracic) to predict apnea and hypopnea events in a 30-second period using the prior 60 seconds' data. We evaluate the performance of the proposed methods for automatically learning the required features and predicting the sleep apnea events on a large dataset containing 48,000 examples from 1,507 subjects. The results show the effectiveness of the proposed convolutional neural network method, which achieved accuracy of 80.78% and F1 score of 80.63%. We also analyse the Markov chain rules and provide an overview of the transitions between apnea and normal events.

We investigate the application of Temporal Convolutional Neural Networks (TCNNs) for solar power forecasting. TCNN is a novel convolutional architecture designed for sequential modelling, which combines causal and dilated convolutions and residual connections. We compare the performance of TCNN with multi-layer feedforward neural networks, and also with recurrent networks, including the state-of-the-art LSTM and GRU recurrent networks. The evaluation is conducted on two Australian datasets containing historical solar and weather data, and weather forecast data for future days. Our results show that TCNN outperformed the other models in terms of accuracy and was able to maintain a longer effective history compared to the recurrent networks. This highlights the potential of convolutional architectures for solar power forecasting tasks.

Tree data occurs in many forms, such as computer programs, chemical molecules, or natural language. Unfortunately, the non-vectorial and discrete nature of trees makes it challenging to construct functions with tree-formed output, complicating tasks such as optimization or time series prediction. Autoencoders address this challenge by mapping trees to a vectorial latent space, where tasks are easier to solve, and then mapping the solution back to a tree structure. However, existing autoencoding approaches for tree data fail to take the specific grammatical structure of tree domains into account and rely on deep learning, thus requiring large training datasets and long training times. In this paper, we propose tree echo state autoencoders (TES-AE), which are guided by a tree grammar and can be trained within seconds by virtue of reservoir computing. In our evaluation on three datasets, we demonstrate that our proposed approach is not only much faster than a state-of-the-art deep learning autoencoding approach (D-VAE) but also has less autoencoding error if little data and time is given.

Personality detection is an important task in psychology, as different personality traits are linked to different behaviours and real-life outcomes. Traditionally it involves filling out lengthy questionnaires, which is time-consuming, and may also be unreliable if respondents do not fully understand the questions or are not willing to honestly answer them. In this article, we propose a framework for objective personality detection that leverages humans’ physiological responses to external stimuli. We exemplify and evaluate the framework in a case study, where we expose subjects to affective image and video stimuli, and capture their physiological responses using non-invasive commercial-grade eye-tracking and skin conductivity sensors. These responses are then processed and used to build a machine learning classifier capable of accurately predicting a wide range of personality traits. We investigate and discuss the performance of various machine learning methods, the most and least accurately predicted traits, and also assess the importance of the different stimuli, features, and physiological signals. Our work demonstrates that personality traits can be accurately detected, suggesting the applicability of the proposed framework for robust personality detection and use by psychology practitioners and researchers, as well as designers of personalised interactive systems.

Personality is an established domain of research in psychology, and individual differences in various traits are linked to a variety of real-life outcomes and behaviours. Personality detection is an intricate task that typically requires humans to fill out lengthy questionnaires assessing specific personality traits. The outcomes of this, however, may be unreliable or biased if the respondents do not fully understand or are not willing to honestly answer the questions. To this end, we propose a framework for objective personality detection that leverages humans' physiological responses to external stimuli. We exemplify and evaluate the framework in a case study, where we expose subjects to affective image and video stimuli, and capture their physiological responses using a commercial-grade eye-tracking sensor. These responses are then processed and fed into a classifier capable of accurately predicting a range of personality traits. Our work yields notably high predictive accuracy, suggesting the applicability of the proposed framework for robust personality detection.

Abstract In this paper, we consider the task of predicting the electricity power generated by photovoltaic solar systems for the next day at half‐hourly intervals. We introduce DL, a deep learning approach based on feed‐forward neural networks for big data time series, which decomposes the forecasting problem into several sub‐problems. We conduct a comprehensive evaluation using 2 years of Australian solar data, evaluating accuracy and training time, and comparing the performance of DL with two other advanced methods based on neural networks and pattern sequence similarity. We investigate the use of multiple data sources (solar power and weather data for the previous days, and weather forecast for the next day) and also study the effect of different historical window sizes. The results show that DL produces competitive accuracy results and scales well, and is thus a highly suitable method for big data environments.

Sleep apnea is a sleep disorder characterized by abnormal breathing patterns during sleep, and affecting 2-4% of the adult population. If left untreated, it increases the risk of heart attack, stroke, diabetes, depression and early death. There are two main types of abnormal breathing events: obstructive apnea and hypopnea. Detecting these events using the traditional machine learning approaches requires extraction and selection of suitable features from several respiratory channels, that are then used as inputs to a classification model. In this study, we present a new approach, based on convolutional neural networks, that automatically combines the raw data of three respiratory channels of polysomnography recordings (nasal airflow, thoracic and abdominal), without feature engineering, and classifies each 30s epoch as containing normal, obstructive apnea or hypopnea events. The evaluation was conducted on a large dataset from 1,507 subjects, containing 23,088 epochs from each of the three classes. We also studied the effectiveness of the individual channels for improving the classification accuracy by testing all channel combinations. Our results showed that the use of nasal airflow, thoracic and abdominal channels with a convolutional neural network was beneficial in detecting sleep apnea events. The combined use of the three channels outperformed all single and pair combinations of channels, achieving accuracy of 83.5%, which is sufficiently high for practical applications.

We investigate the application of convolutional neural networks for energy time series forecasting. In particular, we consider predicting the photovoltaic solar power and electricity load for the next day, from previous solar power and electricity loads. We compare the performance of convolutional neural networks with multilayer perceptron neural networks, which are one of the most popular and successful methods used for these tasks, and also with long short-term memory recurrent neural networks and a persistence baseline. The evaluation is conducted using four solar and electricity time series from three countries. Our results showed that the convolutional and multilayer perceptron neural networks performed similarly in terms of accuracy and training time, and outperformed the other models. This highlights the potential of convolutional neural networks for energy time series forecasting.

Accurate forecasting of the power generated by PhotoVoltaic (PV) systems is needed for the successful integration of solar power into the electricity grid. In this paper, we consider the task of simultaneously forecasting the PV power output for the next day at half-hourly intervals using only previous PV power data. We propose a number of methods for constructing static and dynamic ensembles of neural networks. The static ensembles are based on random sampling and random feature selection, and the dynamic ones adaptively weight the contribution of the ensemble members based on their recent performance. We conduct an evaluation using Australian solar PV data for two years, and compare the results with state-of-the-art single prediction models and classical ensemble models. The results show that the proposed static ensembles are beneficial, achieving higher accuracy than the single neural networks and the other single and ensemble models used for comparison. The dynamic ensemble versions further improve the accuracy, demonstrating the potential of adaptively combining the predictions of ensemble members for accurate solar power forecasting.

Customer behavior modeling is important for businesses in order to understand, attract and retain customers. It is critical that the models are able to track the dynamics of customer behavior over time. We propose FC-CSM, a Customer Segmentation Model based on a Fragmentation-Coagulation process, which can track the evolution of customer segmentation, including the splitting and merging of customer groups. We conduct a case study using transaction data from a major Australian supermarket chain, where we: 1) show that our model achieves high fitness of purchase rate, outperforming models using mixture of Poisson processes; 2) compare the impact of promotions on customers for different products; and 3) track how customer groups evolve over time and how individual customers shift across groups. Our model provides valuable information to stakeholders about the different types of customers, how they change purchase behavior, and which customers are more receptive to promotion campaigns.

We consider the task of predicting the half-hourly solar PhotoVoltaic (PV) power output for the next day from three sources: previous solar power, previous weather data and weather forecasts. We propose DWkNN, a data source weighed nearest neighbor method that considers the importance of the different data sources and learns the best weights for them. We evaluate its performance using Australian PV and weather data for one year. Our results show that DWkNN is more accurate than the standard nearest neighbor and other machine learning, statistical and persistence methods used for comparison. We assess the importance of the three data sources to predict the PV power output and show that the most useful information source is the weather forecast, followed by the previous weather and previous PV power data. The addition of the daily average solar irradiance forecast for the next day, when this information is available, considerably improves the accuracy.

Although solar power is one of the most widely used renewable energy sources, it is highly variable and needs accurate forecasting for its large-scale integration into the electricity grid. We propose WPP, a Weather type Pair Pattern approach, for directly and simultaneously predicting the solar power output for the next day at half-hourly intervals. WPP firstly partitions the days from the training data into clusters based on their weather characteristics and then uses the cluster label of the consecutive days to form pair patterns. A separate prediction model is built for each pair pattern group, which takes as an input the solar power output for the previous day and predicts the one for the next day. The performance of WPP is evaluated using two years of Australian data and compared with a number of state-of-the-art methods and baselines. The results show that WPP is the most accurate method, demonstrating the advantage of using the similarity between the weather characteristics of the consecutive days, in addition to clustering the days and building specialized prediction models for each group.

The supermarkets often use sales promotions to attract customers and create brand loyalty. They would often like to know if their promotions are effective for various customers, so that better timing and more suitable rate can be planned in the future. Given a transaction data set collected by an Australian national supermarket chain, in this paper we conduct a case study aimed at discovering customers' long-term purchase patterns, which may be induced by preference changes, as well as short-term purchase patterns, which may be induced by promotions. Since purchase events of individual customers may be too sparse to model, we propose to discover a number of latent purchase patterns from the data. The latent purchase patterns are modeled via a mixture of non-homogeneous Poisson processes where each Poisson intensity function is composed by long-term and short-term components. Through the case study, 1) we validate that our model can accurately estimate the occurrences of purchase events; 2) we discover easy-to-interpret long-term gradual changes and short-term periodic changes in different customer groups; 3) we identify the customers who are receptive to promotions through the correlation between behavior patterns and the promotions, which is particularly worthwhile for target marketing.

The multichannel nature of EEG and EMG data poses a big challenge to the development of automatic EEG/EMG analysis and classification systems. Due to the “curse of dimensionality” problem, the analysis and classification of several channels may not lead to the desired performance. Accordingly, a number of algorithms have been proposed to identify small “static” subsets of channels that are capable of differentiating between samples of different classes. However, the identification of small subsets of relevant channels may not always be possible, where for certain applications the smaller the number of channels the less chance that sufficient information is provided. We propose in this paper a dynamic channel selection algorithm that identifies a channel (or a subset of channels) for each time segment of the signal that is relevant to the class of that particular time segment. To achieve this, we embraced the dynamic classifier selection methodology, and particularly the multiple classifier behaviour approach. Each EEG/EMG channel can be chosen to represent a certain unseen time segment of the signal based on the performance, or local accuracy, of its nearest neighbours in the set of training time segments. Results obtained using EEG data of a four-class alertness state classification problem reveal that the proposed approach is capable of achieving competitive performance compared to a traditional static channel selection based method. The algorithm also produced very encouraging results when used to classify EMG data collected from nine transradial amputees performing six classes of movements.

We consider the task of forecasting the electricity power generated by a photovoltaic solar system, for the next day at half-hourly intervals. The forecasts are based on previous power output and weather data, and weather prediction for the next day. We present a new approach that forecasts all the power outputs for the next day simultaneously. It builds separate prediction models for different types of days, where these types are determined using clustering of weather patterns. As prediction models it uses ensembles of neural networks, trained to predict the power output for a given day based on the weather data. We evaluate the performance of our approach using Australian photovoltaic solar data for two years. The results showed that our approach obtained MAE=83.90 kW and MRE=6.88%, outperforming four other methods used for comparison.

Accurate forecasting of solar power is needed for the successful integration of solar energy into the electricity grid. In this paper we consider the task of predicting the half-hourly solar photovoltaic power for the next day from previous solar power and weather data. We propose and evaluate several clustering based methods, that group the days based on the weather characteristics and then build a separate prediction model for each cluster using the solar power data. We compare these methods with their non-clustering based counterparts, and also with non-clustering based methods that build a single prediction model for all types of days. We conduct a comprehensive evaluation using Australian data for two years. Our results show that the most accurate prediction model was the clustering based nearest neighbor which uses a vector of half-hourly solar irradiance for the clustering. It achieved MAE=59.81 KW, outperforming all other clustering and non-clustering based methods and baselines.

In this paper we present a new approach for time series forecasting, called Maximum Length Weighted Nearest Neighbor (MLWNN), which combines prediction based on sequence similarity with optimization techniques. MLWNN predicts the 24 hourly electricity loads for the next day, from a time sequence of previously electricity loads up to the current day. We evaluate MLWNN using electricity load data for two years, for three countries (Australia, Portugal and Spain), and compare its performance with three state-of-the-art methods (weighted nearest neighbor, pattern sequence-based forecasting and iterative neural network) and with two baselines. The results show that MLWNN is a promising approach for one day ahead electricity load forecasting.

Forecasting solar power generated from photovoltaic systems at different time intervals is necessary for ensuring reliable and economic operation of the electricity grid. In this paper, we study the application of neural networks for predicting the next day photovoltaic power outputs in 30 minutes intervals from the previous values, without using any exogenous data. We propose three different approaches based on ensembles of neural networks - two non-iterative and one iterative. We evaluate the performance of these approaches using four Australian solar datasets for one year. This includes assessing predictive accuracy, evaluating the benefit of using an ensemble, and comparing performance with two persistence models used as baselines and a prediction model based on support vector regression. The results show that among the three proposed approaches, the iterative approach was the most accurate and it also outperformed all other methods used for comparison.

We present INN, a new approach for predicting the hourly electricity load profile for the next day from a time series of previous electricity loads. It uses an iterative methodology to make the predictions for the 24-hour forecasting horizon. INN combines an efficient mutual information feature selection method with a neural network forecasting algorithm. We evaluate INN using two years of electricity load data for Australia, Portugal and Spain. The results show that it provides accurate predictions, outperforming three state-of-the-art approaches (weighted nearest neighbor, pattern sequence similarity and iterative linear regression), and a number of baselines. INN is also more accurate and efficient than a non-iterative version of the approach. We also found that although the range of load values for the three countries is very different, the load curves show similar patterns, which resulted in more than 90% overlap in the selected lag variables.

Direct quotations are used for opinion mining and information extraction as they have an easy to extract span and they can be attributed to a speaker with high accuracy.However, simply focusing on direct quotations ignores around half of all reported speech, which is in the form of indirect or mixed speech.This work presents the first large-scale experiments in indirect and mixed quotation extraction and attribution.We propose two methods of extracting all quote types from news articles and evaluate them on two large annotated corpora, one of which is a contribution of this work.We further show that direct quotation attribution methods can be successfully applied to indirect and mixed quotation attribution.* *These authors contributed equally to this work.by quotation marks, which makes them easy to extract.However, annotated resources suggest that direct quotations represent only a limited portion of all quotations, i.e., around 30% in the Penn Attribution Relation Corpus (PARC), which covers Wall Street Journal articles, and 52% in the Sydney Morning Herald Corpus (SMHC), with the remainder being indirect (Ex.1c) or mixed (Ex.1b)quotations.Retrieving only direct quotations can miss key content that can change the interpretation of the quotation (Ex.1b) and will entirely miss indirect quotations.

We present PSF-NN, a new approach for time series forecasting. It combines prediction based on sequence similarity with neural networks. PSF-NN first generates predictions using the PSF algorithm that are then refined by the neural network component, which also utilizes additional features. We evaluate the performance of PSF-NN using a time series of hourly electricity demands for the state of New South Wales in Australia for three years. The task is to predict an interval of future values simultaneously, i.e. the 24 demands for the next day, instead of predicting just a single future demand. The results showed that the combined PSF-NN approach provides accurate predictions, outperforming the original PSF algorithm and a number of baselines.

Most of the research in time series is concerned with point forecasting. In this paper we focus on interval forecasting and its application for electricity load prediction. We extend the LUBE method, a neural network-based method for computing prediction intervals. The extended method, called LUBEX, includes an advanced feature selector and an ensemble of neural networks. Its performance is evaluated using Australian electricity load data for one year. The results showed that LUBEX is able to generate high quality prediction intervals, using a very small number of previous lag variables and having acceptable training time requirements. The use of ensemble is shown to be critical for the accuracy of the results.

Web-based catch-up TV has revolutionised watching habits as it provides users the opportunity to watch programs at their preferred time and place, using a variety of devices. With the increasing offer of TV content, there is an emergent need for personalised recommendation solutions, which help users to select programs of interest. In this work, we study the watching patterns of users of an Australian nation-wide catch-up TV service provider and develop a suite of approaches for a catch-up recommendation scenario. We evaluate these approaches using a new large-scale dataset gathered by the Web-based catch-up portal deployed by the provider. The evaluation allows us to compare the performance of several recommenders that address the discovery of both TV programs already watched by users and new programs that users may find relevant.

We present a new approach for electricity load forecasting based on non-decimated multilevel wavelet transform, in combination with two-stage feature selection and machine learning prediction algorithm. The key idea is to decompose the non-stationary and noisy electricity load data into sub-series of different frequencies, analyse and predict them separately. The feature selection integrates autocorrelation and ranking-based methods. We evaluate the predictive performance of our approach using two years of Australian electricity data. The results show that it provides accurate predictions, outperforming exponential smoothing with single and double seasonality, the industry model and all other baselines.

We present new approaches for building yearly and seasonal models for 5-minute ahead electricity load forecasting. They are evaluated using two full years of Australian electricity load data. We first analyze the cyclic nature of the electricity load and show that the autocorrelation function captures these patterns and can be used to extract useful features, as the data is highly linearly correlated. Using the selected feature sets, we then evaluate the predictive performance of four algorithms, representing different prediction paradigms. We found linear regression to be the most accurate and fastest algorithm, outperforming the industry model based on backpropagation neural networks and all baselines. Our results also show that there is no accuracy gain in building models for each season in comparison to building a single yearly model.

We use autocorrelation analysis to extract 6 nested feature sets of previous electricity loads for 5-minite ahead electricity load forecasting. We evaluate their predictive power using Australian electricity data. Our results show that the most important variables for accurate prediction are previous loads from the forecast day, 1, 2 and 7 days ago. By using also load variables from 3 and 6 days ago, we achieved small further improvements. The 3 bigger feature sets (37-51 features) when used with linear regression and support vector regression algorithms, were more accurate than the benchmarks. The overall best prediction model in terms of accuracy and training time was linear regression using the set of 51 features.

We present new approaches for 5-minute ahead electricity load forecasting. They were evaluated on data from the Australian electricity market operator for 2006-2008. After examining the load characteristics using autocorrelation analysis with 4-week sliding window, we selected 51 features. Using this feature set with linear regression and support vector regression we achieved an improvement of 7.56% in the Mean Absolute Percentage Error (MAPE) over the industry model which uses backpropagation neural network. We then investigated the application of a number of methods for further feature subset selection. Using a subset of 38 and 14 of these features with the same algorithms we were able to achieve an improvement of 6.53% and 4.81% in MAPE, respectively, over the industry model.

The reciprocal recommender is a class of recommender system that is important for several tasks where people are both the subjects and objects of the recommendation. Some examples are: job recommendation, mentor-mentee matching, and online dating. Despite the importance of this type of recommender, our work is the first to distinguish it and define its properties. We have implemented RECON, a reciprocal recommender for online dating, and have evaluated it on a large dataset from a major Australian dating website. We investigated the predictive power gained by taking account of reciprocity, finding that it is substantial, for example it improved the success rate of the top ten recommendations from 23% to 42% and also improved the recall at the same time. We also found reciprocity to help with the cold start problem obtaining a success rate of 26% for the top ten recommendations for new users. We discuss the implications of these results for broader uses of our approach for other reciprocal recommenders.

Reciprocal recommender is a class of recommender systems that is important for tasks where people are both the subject and the object of the recommendation; one such task is online dating. We have implemented RECON, a reciprocal recommender for online dating, and we have evaluated it on a major dating website. Results show an improved success rate for recommendations that consider reciprocity in comparison to recommendations that only consider the preferences of the users receiving the recommendations.

In this paper, we present a new approach for very short term electricity load demand forecasting. In particular, we apply support vector regression to predict the load demand every 5 minutes based on historical data from the Australian electricity operator NEMMCO for 2006-2008. The results show that support vector regression is a very promising approach, outperforming backpropagation neural networks, which is the most popular prediction model used by both industry forecasters and researchers. However, it is interesting to note that support vector regression gives similar results to the simpler linear regression and least means squares models. We also discuss the performance of four different feature sets with these prediction models and the application of a correlation-based sub-set feature selection method.

Group work is widespread in education. The growing use of online tools supporting group work generates huge amounts of data. We aim to exploit this data to support mirroring: presenting useful high-level views of information about the group, together with desired patterns characterizing the behavior of strong groups. The goal is to enable the groups and their facilitators to see relevant aspects of the group's operation and provide feedback if these are more likely to be associated with positive or negative outcomes and indicate where the problems are. We explore how useful mirror information can be extracted via a theory-driven approach and a range of clustering and sequential pattern mining. The context is a senior software development project where students use the collaboration tool TRAC. We extract patterns distinguishing the better from the weaker groups and get insights in the success factors. The results point to the importance of leadership and group interaction, and give promising indications if they are occurring. Patterns indicating good individual practices were also identified. We found that some key measures can be mined from early data. The results are promising for advising groups at the start and early identification of effective and poor practices, in time for remediation.

In this paper we propose a new graph-based feature splitting algorithm maxlnd, which creates a balanced split maximizing the independence between the two feature sets. We study the performance of RBF net in a co-training setting with natural, truly independent, random and maxlnd split. The results show that RBF net is successful in a co-training setting, outperforming SVM and NB. Co-training is also found to be sensitive to the trade-off between the dependence of the features within a feature set, and the dependence between the feature sets.

This paper presents a content-adaptive system for streaming reconstructed soccer goal events over networks with bandwidth limited to 1.5Mbps or below. The reconstruction module analyzes a soccer video to produce corresponding panoramic field model with localized motion trajectories to provide tactic analysis for user's better comprehension of the goal events. The transmission module designs 3 schemes for different bandwidth conditions (150Kbps-1.5Mbps), (56-150Kbps), and (<56Kbps). Each scheme can choose dynamically the video content, viewing quality and where the reconstruction happens based on the network conditions and computational power of connected devices. We demonstrate the effectiveness and scalability of our system in experimental results of adaptive transmission two soccer goal events under three typical bandwidth conditions.

Recently, we proposed a new meta-learning approach based on landmarking. This approach, which utilises a new set of criteria for selecting landmarkers, generates a set of landmarkers that are each functions over the performance over subsets of the candidate algorithms being landmarked. In this paper, we experiment with three heuristics based on correlativity and efficiency. With each heuristic, the landmarkers generated using linear regression are able to estimate accuracy well, even when only utilising a small fraction of the given algorithms. The results also show that the heuristic in which efficiencies are estimated via 1-nearest neighbour outperformed the other heuristics.

Landmarking is a recent and promising metalearning strategy, which defines meta-features that are themselves efficient learning algorithms. However, the choice of landmarkers is made in an ad hoc manner. In this paper, we propose a new perspective and set of criteria for landmarkers. With these, we introduce a landmarker generation algorithm, which creates a set of landmarkers that each utilise subsets of the algorithms being landmarked. The experiments show that the landmarkers formed, when used with linear regression, are able to estimate accuracy well, even when utilising a small fraction of the given algorithms. 1.

We present a new approach for video summarization and browsing of MPEG-2 compressed video based on the growing cell structures (GCS) neural algorithms. It first applies GCS to select keyframes for each shot and then clusters them using TreeGCS to form a hierarchical view of the video for efficient browsing. The keyframe selection is based on histogram features of the dc-images for T frames. It captures well the video content and outperforms two other approaches. The main advantage of the TreeGCS module is the ability to form dynamically a flexible hierarchy depending on the video content.

When dealing with information overload from the Internet, such as the classification of Web pages and the filtering of email spam, a new technique called co-training has been shown to be a promising approach to help build more accurate classifiers. Co-training allows classifiers to learn with fewer labelled documents by taking advantage of the more abundant unclassified documents. However, conventional co-training requires the dataset to be described by two disjoint and natural feature sets that are sufficiently redundant. In many practical situations, it is not intuitively obvious how to obtain two natural feature sets. This paper shows that when only a single natural feature set is used, the performance of co-training is beneficial in the application of email classification.

We present a neural network based system for automated e-mail filing into folders and anti-spam filtering. The experiments show that it is more accurate than several other techniques. We also investigate the effects of various feature selection, weighting and normalization methods, and also the portability of the anti-spam filter across different users.

We present INTIMATE, a Web-based movie recommender that makes suggestions by using text categorization to learn from movie synopses. The performance of various feature representations, feature selectors, feature weighting mechanisms and classifiers is evaluated and discussed. INTIMATE was also compared with a feature-based movie recommender. The results show that the text-based approach outperforms the feature-based if the ratio of the number of user ratings to the vocabulary size is high.

Reports an application of an evolving fuzzy neural network (EFuNN) for camera operations recognition. EFuNN features one-pass learning, dynamical growing and shrinking architecture and ability to accommodate new knowledge without the need to retrain the network on both the original and new data. The network learns from pre-classified examples in the form of motion vector patterns, extracted from MPEG compressed video, in order to distinguish between six classes: static, panning, zooming, object motion, tracking and dissolve. The performance of EFuNN is compared with LVQ and the results are discussed. In addition, the impact of the number of membership functions and the contribution of the rule node aggregation are analyzed.