Recommender systems are widely used in electronic commerce, social media and online streaming services to provide personalized recommendations to the users by exploiting past ratings and interactions. This paper considers the security aspect with quick and accurate detection of attacks by observing the newly created profiles sequentially to prevent the damage which may be incurred by the injection of new profiles with dishonest ratings. The proposed framework consists of a latent variable model, which is trained by a variational EM algorithm, followed by a sequential detection algorithm. The latent variable model generates homogeneous representations of the users given their rating history and mixed data-type attributes such as age and gender. The representations are then exploited to generate univariate statistics to be efficiently used in a CUSUM-like sequential detection algorithm that can quickly detect persistent attacks while maintaining low false alarm rates. We apply our proposed framework to three different real-world datasets and exhibit superior performance in comparison to the existing baseline algorithms for both attack profile and sequential detection. Furthermore, we demonstrate robustness to different attack strategies and configurations.

Deep learning found its initial footing in supervised applications such as image and voice recognition successes of which were followed by deep generative models across similar domains. In recent years, researchers have proposed creative learning representations to utilize the unparalleled generalization capabilities of such structures for unsupervised applications commonly called deep clustering. This paper presents a comprehensive analysis of popular deep clustering architectures including deep autoencoders and convolutional autoencoders to study how network topology, hyperparameters and clustering coefficients impact accuracy. Three popular benchmark datasets are used including MNIST, CIFAR10 and SVHN to ensure data independent results. In total, 20 different pairings of topologies and clustering coefficients are used for both the standard and convolutional autoencoder architectures across three different datasets for a joint analysis of 120 unique combinations with sufficient repetitive testing for statistical significance. The results suggest that there is a general optimum when it comes to choosing the coding layer (latent dimension) size which is correlated to an extent with the complexity of the dataset. Moreover, for image datasets, when color makes a meaningful contribution to the identity of the observation, it also helps improve the subsequent deep clustering performance.

Aerial deployment of access points (APs) to network infrastructure has been shown to be an effective technological solution for rapidly supplying networking services in diverse scenarios, particularly post-disaster networking after major hurricane, tsunami, and earthquake events, or for the purposes of homeland security. A straightforward yet helpful method for deployment of APs in this emerging technology is to consider the problem in two 2D planes, one for users and another for APs. However, considering the need for effectively connecting low-power user devices that may suffer from severe fading due to the blocking of line-of-sight(LoS) transmissions, genuine 3D deployment is very desirable, even though a simpler form of it had been mathematically an open problem until 2017 for 400 years. To meet the engineering requirements, we first leverage 3D random geometric graphs (RGG) to develop the joint optimization of 3D AP placement with a given number of APs for proper coverage. We further propose a computationally-efficient iterative algorithm to effectively deploy the adequate number of APs in 3D space and respond to dynamics of such wireless networks. Our simulations demonstrate the superior performance of this genuine 3D placement of APs over traditional approaches when considering connections for user devices, particularly for larger number of APs in rapid deployment.

Malicious profiles have been a credible threat to collaborative recommender systems. Attackers provide fake item ratings to systematically manipulate the platform. Attack detection algorithms can identify and remove such users by observing rating distributions. In this study, we aim to use the user attributes as an additional information source to improve the accuracy and speed of attack detection. We propose a probabilistic factorization model which can embed mixed data type user attributes and observed ratings into a latent space to generate anomaly statistics for new users. To identify the persistent outliers in the system, we also propose a sequential attack detection algorithm to enable quick and accurate detection based on the probabilistic model learned from genuine users. The proposed model demonstrates significant improvements in both accuracy and speed when compared to baseline algorithms on a popular benchmark dataset.

In this paper, we propose a novel multimodal framework for matrix completion where the missing values are accumulatively estimated with feature-based neural networks. Specifically, for each individual feature, a different model is trained to predict the missing values in the observations using the remaining features as input to provide an initial estimate. These estimates are then used to initiate the next round of model training to iteratively converge to the final prediction of the missing value. The weight parameters of the networks are propagated through these accumulative iterations which leads to a computationally efficient algorithm where training times become progressively shorter with each round. Unlike traditional algorithms relying on matrix factorization, separating the model building and exploiting steps also enables more effective deployment in online applications. Results show that the proposed algorithm outperforms prior work in 80% of the test scenarios when compared to four universally accepted methods on a combination of different datasets and missing data ratios.

Passive ultra-high frequency (UHF) radio frequency identification (RFID) systems have gained immense popularity due to groundbreaking advances in wireless devices and technologies for their wide-scale industrial applications in inventory tracking and management. In this study, we explore the potential of passive RFID systems for indoor localization by developing a grid-based experimental framework using two standard and easily measurable performance metrics: received signal strength indicator (RSSI) and tag read count (TRC). We create scenarios imitating real life challenges such as placing metal objects and other RFID tags in two different read fields (symmetric and asymmetric) to analyze their impacts on location accuracy. We study the prediction potential of RSSI and TRC both independently and collaboratively. In the end, we demonstrate that both signal metrics can be used for localization with sufficient accuracy whereas the best performance is obtained when both metrics are used together for prediction on an artificial neural network especially for more challenging scenarios. Experimental results show an average error of as low as 0.286 (where consecutive grid distance is defined as unity) which satisfies the grid-based localization benchmark of less than 0.5.

Support vector machines (SVM), originally introduced as powerful binary classifiers, can also be used for multi-class recognition with the help of creative meta-learning strategies such as commonly used one-vs-rest, one-vs-one and majority voting. In this paper, we explore the potential of creating informed nested dichotomies based on clustering pseudo-labels and probability estimates generated a priori through separate supervised training. Across six diverse public datasets, we found that in majority of the cases the final tree structure is unbalanced with one-off elimination of a single class in each branch. When compared to traditional multi-class SVM topologies, the proposed method shows no statistically significant change in accuracy. However, the on-demand testing time is reduced quite substantially with a range of 2-to-25-fold decrease in the computational load compared to baseline classifiers across different datasets. Finally, the proposed algorithm can be adjusted for class frequencies in imbalanced datasets to improve testing time even further for applications such as letter recognition.

As mobile and personal health devices gain in popularity, increasing amounts of data is collected via their embedded sensors such as heart rate monitors and accelerometers. Data analytics and more specifically machine learning algorithms can transform this data into actionable information to improve personal healthcare and quality of life. The main objective of this study is to develop an algorithmic classification framework using feed-forward multilayer perceptrons and statistically rich spectrotemporal features to recognize daily activities based on 3-axis acceleration data. A multitude of MLP topologies and setups, such as different numbers and sizes of hidden layers, supervised output structuring, etc. are tested to comprehensively analyze the clustering capabilities of the artificial neural network for a wide-range of settings. In addition, the contribution of subset of features to classification accuracy is studied to identify respective information potentials and further improve accuracy. Publicly available wrist-worn accelerometer dataset from University of California Irvine's machine learning repository is used for fair comparison with the most recent literature published using the same dataset. Results indicate a significant improvement in recognition rate where the overall accuracy over seven selected activity classes is 91% compared to 54% of the latest publication using the same dataset.

One of the most critical components of engineering a machine learning algorithm for a live application is robust performance assessment prior to its implementation. Cross-validation is used to forecast a specific algorithm's classification or prediction accuracy on new input data given a finite dataset for training and testing the algorithm. Two most well known cross-validation techniques, random subsampling (RSS) and K-fold, are used to generalize the assessment results of machine learning algorithms in a non-exhaustive random manner. In this work we first show that for an inertia based activity recognition problem where data is collected from different users of a wrist-worn wireless accelerometer, random partitioning of the data, regardless of cross-validation technique, results in statistically similar average accuracies for a standard feed-forward neural network classifier. We propose a novel source-aware partitioning technique where samples from specific users are completely left out of the training/validation sets in rotation. The average error for the proposed cross-validation method is significantly higher with lower standard variation, which is a major indicator of cross-validation robustness. Approximately 30% increase in average error rate implies that source-aware cross validation could be a better indication of live algorithm performance where test data statistics would be significantly different than training data due to source (or user)-sensitive nature of process data.

A biologically plausible algorithm is proposed for phoneme recognition that makes use of spikes for computation. The prototype system is demonstrated on voiced phonemes and shows competitive performance with state-of-the-art systems on a vowel dataset in the presence of noise. Using novel phase locking cues, the algorithm performs surprisingly well down to SNR values as low as 5 dB, where the performance of the baseline algorithm used for comparison drops considerably. These results suggest, not only a possible explanation for the extreme noise robustness of human recognition, but also a novel technique for information processing for speech signals with computers.

The insufficiency of using only second-order statistics and premise of exploiting higher order statistics of the data has been well understood, and more advanced objectives including higher order statistics, especially those stemming from information theory, such as error entropy minimization, are now being studied and applied in many contexts of machine learning and signal processing. In the adaptive system training context, the main drawback of utilizing output error entropy as compared to correlation-estimation-based second-order statistics is the computational load of the entropy estimation, which is usually obtained via a plug-in kernel estimator. Sample-spacing estimates offer computationally inexpensive entropy estimators; however, resulting estimates are not differentiable, hence, not suitable for gradient-based adaptation. In this brief paper, we propose a nonparametric entropy estimator that captures the desirable properties of both approaches. The resulting estimator yields continuously differentiable estimates with a computational complexity at the order of those of the sample-spacing techniques. The proposed estimator is compared with the kernel density estimation (KDE)-based entropy estimator in the supervised neural network training framework with computation time and performance comparisons.

We propose a duplex theory of spike coding in the early stages of the auditory system based on the intensity and noise levels of the acoustic stimuli. According to this concept, at low intensity levels, where auditory nerve firings cannot generate a high enough synchrony among neuron ensembles, rate coding is more likely favored against phase-locking via synchrony coding. To the contrary, at conversational intensity levels, phase synchrony coding is preferred due to its superior and highly noise robust performance. The theory is supported by both evidence from biology, as well as from experimental simulations using biologically plausible models of the entire processing chain from spike generation to recognition.

We propose a noise robust feature extraction technique for speech signals using phase synchrony. The front-end employs a psychoacoustic cochlea model with inner hair cells to transform speech into a parallel stream of spike trains as observed in the auditory nerve fibers. The degree of phase synchrony among nerve fibers with similar characteristic frequencies is calculated to yield a feature vector which shows little degradation in response to increasing levels of noise. As a benchmark, the feature set is used in a biologically plausible model with a spike-based, liquid state machine classifier for a simple acoustic classification task. Though applied to a simplified domain, the results indicate a superior performance when compared to a conventional speech recognition system, especially at very low signal-to-noise ratios.