Nowadays, many classification algorithms have been applied to various industries to help them work out their problems met in real-life scenarios. However, in many binary classification tasks, samples in the minority class only make up a small part of all instances, which leads to the datasets we get usually suffer from high imbalance ratio. Existing models sometimes treat minority classes as noise or ignore them as outliers encountering data skewing. In order to solve this problem, we propose a bagging ensemble learning framework $ASE$ (Anomaly Scoring Based Ensemble Learning). This framework has a scoring system based on anomaly detection algorithms which can guide the resampling strategy by divided samples in the majority class into subspaces. Then specific number of instances will be under-sampled from each subspace to construct subsets by combining with the minority class. And we calculate the weights of base classifiers trained by the subsets according to the classification result of the anomaly detection model and the statistics of the subspaces. Experiments have been conducted which show that our ensemble learning model can dramatically improve the performance of base classifiers and is more efficient than other existing methods under a wide range of imbalance ratio, data scale and data dimension. $ASE$ can be combined with various classifiers and every part of our framework has been proved to be reasonable and necessary.

Understanding model predictions is critical in healthcare, to facilitate rapid verification of model correctness and to guard against use of models that exploit confounding variables. We introduce the challenging new task of explainable multiple abnormality classification in volumetric medical images, in which a model must indicate the regions used to predict each abnormality. To solve this task, we propose a multiple instance learning convolutional neural network, AxialNet, that allows identification of top slices for each abnormality. Next we incorporate HiResCAM, an attention mechanism, to identify sub-slice regions. We prove that for AxialNet, HiResCAM explanations are guaranteed to reflect the locations the model used, unlike Grad-CAM which sometimes highlights irrelevant locations. Armed with a model that produces faithful explanations, we then aim to improve the model's learning through a novel mask loss that leverages HiResCAM and 3D allowed regions to encourage the model to predict abnormalities based only on the organs in which those abnormalities appear. The 3D allowed regions are obtained automatically through a new approach, PARTITION, that combines location information extracted from radiology reports with organ segmentation maps obtained through morphological image processing. Overall, we propose the first model for explainable multi-abnormality prediction in volumetric medical images, and then use the mask loss to achieve a 33% improvement in organ localization of multiple abnormalities in the RAD-ChestCT data set of 36,316 scans, representing the state of the art.

Class imbalance is a frequently occurring scenario in classification tasks. Learning from imbalanced data poses a major challenge, which has instigated a lot of research in this area. Data preprocessing using sampling techniques is a standard approach to deal with the imbalance present in the data. Since standard classification algorithms do not perform well on imbalanced data, the dataset needs to be adequately balanced before training. This can be accomplished by oversampling the minority class or undersampling the majority class. In this study, a novel hybrid sampling algorithm has been proposed. To overcome the limitations of the sampling techniques while ensuring the quality of the retained sampled dataset, a sophisticated framework has been developed to properly combine three different sampling techniques. Neighborhood Cleaning rule is first applied to reduce the imbalance. Random undersampling is then strategically coupled with the SMOTE algorithm to obtain an optimal balance in the dataset. This proposed hybrid methodology, termed "SMOTE-RUS-NC", has been compared with other state-of-the-art sampling techniques. The strategy is further incorporated into the ensemble learning framework to obtain a more robust classification algorithm, termed "SRN-BRF". Rigorous experimentation has been conducted on 26 imbalanced datasets with varying degrees of imbalance. In virtually all datasets, the proposed two algorithms outperformed existing sampling strategies, in many cases by a substantial margin. Especially in highly imbalanced datasets where popular sampling techniques failed utterly, they achieved unparalleled performance. The superior results obtained demonstrate the efficacy of the proposed models and their potential to be powerful sampling algorithms in imbalanced domain.

We propose a hybrid machine learning architecture that simultaneously employs multiple deep learning models analyzing contextual and behavioral characteristics of Windows portable executable, producing a final prediction based on a decision from the meta-model. The detection heuristic in contemporary machine learning Windows malware classifiers is typically based on the static properties of the sample since dynamic analysis through virtualization is challenging for vast quantities of samples. To surpass this limitation, we employ a Windows kernel emulation that allows the acquisition of behavioral patterns across large corpora with minimal temporal and computational costs. We partner with a security vendor for a collection of more than 100k int-the-wild samples that resemble the contemporary threat landscape, containing raw PE files and filepaths of applications at the moment of execution. The acquired dataset is at least ten folds larger than reported in related works on behavioral malware analysis. Files in the training dataset are labeled by a professional threat intelligence team, utilizing manual and automated reverse engineering tools. We estimate the hybrid classifier's operational utility by collecting an out-of-sample test set three months later from the acquisition of the training set. We report an improved detection rate, above the capabilities of the current state-of-the-art model, especially under low false-positive requirements. Additionally, we uncover a meta-model's ability to identify malicious activity in validation and test sets even if none of the individual models express enough confidence to mark the sample as malevolent. We conclude that the meta-model can learn patterns typical to malicious samples from representation combinations produced by different analysis techniques. We publicly release pre-trained models and anonymized dataset of emulation reports.

Credit cards play an exploding role in modern economies. Its popularity and ubiquity have created a fertile ground for fraud, assisted by the cross boarder reach and instantaneous confirmation. While transactions are growing, the fraud percentages are also on the rise as well as the true cost of a dollar fraud. Volume of transactions, uniqueness of frauds and ingenuity of the fraudster are main challenges in detecting frauds. The advent of machine learning, artificial intelligence and big data has opened up new tools in the fight against frauds. Given past transactions, a machine learning algorithm has the ability to 'learn' infinitely complex characteristics in order to identify frauds in real-time, surpassing the best human investigators. However, the developments in fraud detection algorithms has been challenging and slow due the massively unbalanced nature of fraud data, absence of benchmarks and standard evaluation metrics to identify better performing classifiers, lack of sharing and disclosure of research findings and the difficulties in getting access to confidential transaction data for research. This work investigates the properties of typical massively imbalanced fraud data sets, their availability, suitability for research use while exploring the widely varying nature of fraud distributions. Furthermore, we show how human annotation errors compound with machine classification errors. We also carry out experiments to determine the effect of PCA obfuscation (as a means of disseminating sensitive transaction data for research and machine learning) on algorithmic performance of classifiers and show that while PCA does not significantly degrade performance, care should be taken to use the appropriate principle component size (dimensions) to avoid overfitting.

We achieved an F1 score (harmonic mean of precision and recall) of 0.68, which is quite decent considering the limited size of our dataset. My role was related to hyperparameter optimization for the LRCN model, wherein I experimented with different values of the learning rate, dropout, and regularization techniques and how they impacted the results of our model. One important take from the entire experience was how teamwork is crucial to produce an efficient output. The internship was rigorous, with early morning lectures and late night team meetings, but I learned a lot and had fun in the process!

For the first time, we propose using a multiple instance learning based convolution-free transformer model, called Multiple Instance Neuroimage Transformer (MINiT), for the classification of T1-weighted (T1w) MRIs. We first present several variants of transformer models adopted for neuroimages. These models extract non-overlapping 3D blocks from the input volume and perform multi-headed self-attention on a sequence of their linear projections. MINiT, on the other hand, treats each of the non-overlapping 3D blocks of the input MRI as its own instance, splitting it further into non-overlapping 3D patches, on which multi-headed self-attention is computed. As a proof-of-concept, we evaluate the efficacy of our model by training it to identify sex from T1w-MRIs of two public datasets: Adolescent Brain Cognitive Development (ABCD) and the National Consortium on Alcohol and Neurodevelopment in Adolescence (NCANDA). The learned attention maps highlight voxels contributing to identifying sex differences in brain morphometry.

We apply post-processing to the class probability distribution outputs of audio event classification models and employ reinforcement learning to jointly discover the optimal parameters for various stages of a post-processing stack, such as the classification thresholds and the kernel sizes of median filtering algorithms used to smooth out model predictions. To achieve this we define a reinforcement learning environment where: 1) a state is the class probability distribution provided by the model for a given audio sample, 2) an action is the choice of a candidate optimal value for each parameter of the post-processing stack, 3) the reward is based on the classification accuracy metric we aim to optimize, which is the audio event-based macro F1-score in our case. We apply our post-processing to the class probability distribution outputs of two audio event classification models submitted to the DCASE Task4 2020 challenge. We find that by using reinforcement learning to discover the optimal per-class parameters for the post-processing stack that is applied to the outputs of audio event classification models, we can improve the audio event-based macro F1-score (the main metric used in the DCASE challenge to compare audio event classification accuracy) by 4-5% compared to using the same post-processing stack with manually tuned parameters.

Graph embedding techniques are a staple of modern graph learning research. When using embeddings for downstream tasks such as classification, information about their stability and robustness, i.e., their susceptibility to sources of noise, stochastic effects, or specific parameter choices, becomes increasingly important. As one of the most prominent graph embedding schemes, we focus on node2vec and analyse its embedding quality from multiple perspectives. Our findings indicate that embedding quality is unstable with respect to parameter choices, and we propose strategies to remedy this in practice.

Popularity bias is the idea that a recommender system will unduly favor popular artists when recommending artists to users. As such, they may contribute to a winner-take-all marketplace in which a small number of artists receive nearly all of the attention, while similarly meritorious artists are unlikely to be discovered. In this paper, we attempt to measure popularity bias in three state-of-art recommender system models (e.g., SLIM, Multi-VAE, WRMF) and on three commercial music streaming services (Spotify, Amazon Music, YouTube). We find that the most accurate model (SLIM) also has the most popularity bias while less accurate models have less popularity bias. We also find no evidence of popularity bias in the commercial recommendations based on a simulated user experiment.