Hyun Ah Song

Hierarchical Representation Using NMF

In this paper, we propose a representation model that demonstrates hierarchical feature learning using nsNMF. We stack simple unit algorithm into several layers to take step-by-step approach in learning. By utilizing NMF as unit algorithm, our proposed network provides intuitive understanding of the feature development process. It is able to represent the underlying structure of feature hierarchies present in complex data in intuitively understandable manner. Experiments with document data successfully discovered feature hierarchies of concepts in data. We also observed that proposed method results in much better classification and reconstruction performance, especially for small number of features.

A new discriminant NMF algorithm and its application to the extraction of subtle emotional differences in speech

In this study we propose a new feature extraction algorithm, dNMF (discriminant non-negative matrix factorization), to learn subtle class-related differences while maintaining an accurate generative capability. In addition to the minimum representation error for the standard NMF (non-negative matrix factorization) algorithm, the dNMF algorithm also results in higher between-class variance for discriminant power. The multiplicative NMF learning algorithm has been modified to cope with this additional constraint. The cost function was carefully designed so that the extraction of feature coefficients from a single testing pattern with pre-trained feature vectors resulted in a quadratic convex optimization problem in non-negative space for uniqueness. It also resolves issues related to the previous discriminant NMF algorithms. The developed dNMF algorithm has been applied to the emotion recognition task for speech, where it needs to emphasize the emotional differences while de-emphasizing the dominant phonetic components. The dNMF algorithm successfully extracted subtle emotional differences, demonstrated much better recognition performance and showed a smaller representation error from an emotional speech database.

Hierarchical feature extraction by multi-layer non-negative matrix factorization network for classification task

In this paper, we propose multi-layer non-negative matrix factorization (NMF) network for classification task, which provides intuitively understandable hierarchical feature learning process. The layer-by-layer learning strategy was adopted through stacked NMF layers, which enforced non-negativity of both features and their coefficients. With the non-negativity constraint, the learning process revealed latent feature hierarchies in the complex data in intuitively understandable manner. The multi-layer NMF networks was investigated for classification task by studying various network architectures and nonlinear functions. The proposed multilayer NMF network was applied to document classification task, and demonstrated that our proposed multi-layer NMF network resulted in much better classification performance compared to single-layered network, even with the small number of features. Also, through intuitive learning process, the underlying structure of feature hierarchies was revealed for the complex document data.

PowerCast: Mining and Forecasting Power Grid Sequences

What will be the power consumption of our institution at 8am for the upcoming days? What will happen to the power consumption of a small factory, if it wants to double (or half) its production? Technologies associated with the smart electrical grid are needed. Central to this process are algorithms that accurately model electrical load behavior, and forecast future electric power demand. However, existing power load models fail to accurately represent electrical load behavior in the grid. In this paper, we propose PowerCast, a novel domain-aware approach for forecasting the electrical power demand, by carefully incorporating domain knowledge. Our contributions are as follows: 1. Infusion of domain expert knowledge: We represent the time sequences using an equivalent circuit model, the “BIG” model, which allows for an intuitive interpretation of the power load, as the BIG model is derived from physics-based first principles. 2. Forecasting of the power load: Our PowerCast uses the BIG model, and provides (a) accurate prediction in multi-step-ahead forecasting, and (b) extrapolations, under what-if scenarios, such as variation in the demand (say, due to increase in the count of people on campus, or a decision to half the production in our factory etc.) 3. Anomaly detection: PowerCast can spot and, even explain, anomalies in the given time sequences. The experimental results based on two real world datasets of up to three weeks duration, demonstrate that PowerCast is able to forecast several steps ahead, with 59% error reduction, compared to the competitors. Moreover, it is fast, and scales linearly with the duration of the sequences.

Graph-Based Fraud Detection in the Face of Camouflage

Given a bipartite graph of users and the products that they review, or followers and followees, how can we detect fake reviews or follows? Existing fraud detection methods (spectral, etc.) try to identify dense subgraphs of nodes that are sparsely connected to the remaining graph. Fraudsters can evade these methods using camouflage, by adding reviews or follows with honest targets so that they look “normal.” Even worse, some fraudsters use hijacked accounts from honest users, and then the camouflage is indeed organic. Our focus is to spot fraudsters in the presence of camouflage or hijacked accounts. We propose FRAUDAR, an algorithm that (a) is camouflage resistant, (b) provides upper bounds on the effectiveness of fraudsters, and (c) is effective in real-world data. Experimental results under various attacks show that FRAUDAR outperforms the top competitor in accuracy of detecting both camouflaged and non-camouflaged fraud. Additionally, in real-world experiments with a Twitter follower--followee graph of 1.47 billion edges, FRAUDAR successfully detected a subgraph of more than 4, 000 detected accounts, of which a majority had tweets showing that they used follower-buying services.

HomeRun: Scalable Sparse-Spectrum Reconstruction of Aggregated Historical Data.

Recovering a time sequence of events from multiple aggregated and possibly overlapping reports is a major challenge in historical data fusion. The goal is to reconstruct a higher resolution event equence from a mixture of lower resolution samples as accurately as possible. For example, we may aim to disaggregate overlapping monthly counts of people infected with measles into weekly counts. In this paper, we propose a novel data disaggregation method, called HOMERUN, that exploits an alternative representation of the sequence and finds the spectrum of the target sequence. More specifically, we formulate the problem as so-called basis pursuit using the Discrete Cosine Transform (DCT) as a sparsifying dictionary and impose non-negativity and smoothness constraints. HOMERUN utilizes the energy compaction feature of the DCT by finding the sparsest spectral representation of the target sequence that contains the largest (most important) coefficients. We leverage the Alternating Direction Method of Multipliers to solve the resulting optimization problem with scalable and memory efficient steps. Experiments using real epidemiological data show that our method considerably outperforms the state-of-the-art techniques, especially when the DCT of the sequence has a high degree of energy compaction.

TensorCast: forecasting and mining with coupled tensors

Given an heterogeneous social network, can we forecast its future? Can we predict who will start using a given hashtag on twitter? Can we leverage side information, such as who retweets or follows whom, to improve our membership forecasts? We present TensorCast, a novel method that forecasts time-evolving networks more accurately than current state-of-the-art methods by incorporating multiple data sources in coupled tensors. TensorCast is (a) scalable, being linearithmic on the number of connections; (b) effective, achieving over 20% improved precision on top-1000 forecasts of community members; (c) general, being applicable to data sources with different structure. We run our method on multiple real-world networks, including DBLP, epidemiology data, power grid data, and a Twitter temporal network with over 310 million nonzeros, where we predict the evolution of the activity of the use of political hashtags.

Noise-Robust Detection of Symmetric Axes by Self-Correcting Artificial Neural Network

Perception of symmetric image patterns is one of the important stages in visual information processing. However, local interference of the input image disturbs the detection of symmetry in artificial neural network based models. In this paper, we propose a noise-robust neural network model that can correct asymmetric corruptions and returns clear symmetry axes. For efficient detection of bilateral symmetry as well as asymmetry correction, our network adopts directional blurring filters. The filter responses are fed to oscillatory neurons for line extraction, which serializes the activation of multiple symmetry axes. Given an activated symmetry axis, the network estimates the difference of counterparts to generate a masking filter that covers the asymmetric parts. The network reconstructs the ideal mirror-symmetric image with complete symmetry axes by self-correction of corruptions. Through simulations on corrupted images, we verify that our network is superior to Fukushima’s symmetry detection network. Our network successfully presents biologically plausible and robust symmetry perception mechanism.

Self-correcting symmetry detection network

In this paper, we propose a symmetry axis detection network that can correct asymmetric parts by itself. Our network compares directional blurring of omnidirectional image edges, which plays a significant role in asymmetry detection and correction. The output layer consists of oscillatory neurons, which activates symmetry axes one by one. Given activated symmetry axis, network estimates the difference of image edges and generates a masking filter to cover the asymmetric parts. The network reconstructs ideal mirror-symmetric image with complete symmetry axes by self-correction. Our network models flexible symmetry perception of high-level cognitive function of human brain.

Enhanced discrimination of face orientation based on gabor filters

In general, a face analysis relies on the face orientation; therefore, face orientation discrimination is very important for interpreting the situation of people in an image. In this paper, we propose an enhanced approach that is robust to the unwanted variation of the image such as illumination, size of faces, and conditions of picture taken. In addition to the conventional algorithm (Principal Component Analysis and Independent Component Analysis), we imposed the Gabor kernels and Fourier Transform to improve the robustness of the proposed approach. The experimental results validate the effectiveness of the proposed algorithm for five kinds of face orientation (front, quarter left, perfect left, quarter right, and perfect right side of faces). In real application, the proposed algorithm will enable a Human-Computer Interface (HCI) system to understand the image better by extracting reliable information of face orientation.