Theses supervised by Prof. Dr. Süleyman Serdar Kozat
12 theses · İhsan Doğramacı Bilkent University
Durum uzayı yaklaşımı kullanılarak doğrusal olmayan zaman serisi regresyonu için gradyan artırıcı karar ağaçları ve sarımax modellerinin yeni birleşik optimizasyonu
We investigate nonlinear regression or forecasting in an online setting and propose an end-to-end ensemble model that addresses two key challenges: the trade-off between linear and nonlinear modeling, and the disjoint nature of the optimization process. The proposed architecture seamlessly integrates a linear time series model and a nonlinear soft decision tree-based model within a unified structure. Specifically, the architecture employs a Seasonal Autoregressive Integrated Moving Average with Exogenous Regressors (SARIMAX) model to capture key linear patterns in time series data, such as seasonality and trend, alongside a soft gradient boosting decision tree (SGBDT), which is adept at modeling nonlinear dependencies and extracting features directly from raw inputs. Different from existing state-of-the-art hybrid approaches that typically rely on sequential and disjoint training strategies, for the first time in the literature, we introduce a jointly optimized hybrid model in which both the linear and nonlinear components are trained simultaneously. This is achieved through a novel formulation based on state-space representations, which allows the integration of the ARMA-based SARIMAX model and the SGBDT into a single state-space framework. We employ the Extended Kalman Filter (EKF), a nonlinear optimization technique, to efficiently perform the joint training process by leveraging derived state transition and measurement equations. Furthermore, the architecture is modular and extensible, allowing for the substitution or addition of alternative nonlinear models and other linear time series techniques, such as Exponential Smoothing (ETS), as long as state representations are obtained. The optimization component is also flexible, enabling the replacement of the EKF with other techniques such as the Unscented Kalman Filter (UKF) or Particle Filters (PF). Through extensive experiments on well-known real-world datasets, the proposed approach demonstrates superior performance. We provide the source code as a publicly available repository to support further research and ensure reproducibility.
Gradyan artırmalı makinelerle doğrusal olmayan regresyon için çoklu sekanslı yeni bir optimizasyon çerçevesi
Gradient Boosting Machines (GBMs) consistently achieve state-of-the-art performance on a wide range of machine learning applications, particularly for problems tabular data. However, their underlying optimization mechanism largely relies on a greedy form of functional gradient descent. This classical approach, while effective, can be path-dependent under noise and prone to locally optimal updates that yield globally suboptimal models. To address these fundamental limitations, we propose a multi-sequence framework that integrates principles from modern optimization directly into the gradient boosting process.Unlike prior work that attempts direct adaptations of vector-based optimizers such as Nesterov's accelerated gradient, our method employs a decoupled architecture inspired by stochastic primal averaging (SPA). This architecture provides a stable foundation upon which we build a series of adaptive target updaters that translate the mechanics of modern optimizers such as Adam into principled learning signals for weak learners. We provide a theoretical analysis, including a linear convergence proof for our base model under standard smoothness and strong convexity assumptions. Our experimental results demonstrate significant improvements in stability and accuracy over LightGBM and other baselines, proving our framework's ability to offer a new level of control over the optimization trajectory while preserving the simplicity and modularity that make GBMs practical, leading to more robust and powerful gradient boosting models.
Düzgün olmayan şekilde örneklenmiş sıralı verinin işlenmesi
We study classi cation and regression for variable length sequential data, which is either non-uniformly sampled or contains missing samples. In most sequential data processing studies, one considers data sequence is uniformly sampled and complete, i.e., does not contain missing input values. However, non-uniformly sampled sequences and the missing data problem appear in a wide range of elds such as medical imaging and nancial data. To resolve these problems, certain preprocessing techniques, statistical assumptions and imputation methods are usually employed. However, these approaches su er since the statistical assumptions do not hold in general and the imputation of arti cially generated and unrelated inputs deteriorate the model. To mitigate these problems, in chapter 2, we introduce a novel Long Short-Term Memory (LSTM) architecture. In particular, we extend the classical LSTM network with additional time gates, which incorporate the time information as a nonlinear scaling factor on the conventional gates. We also provide forward pass and backward pass update equations for the proposed LSTM architecture. We show that our approach is superior to the classical LSTM architecture, when there is correlation between time samples. In chapter 3, we investigate regression for variable length sequential data containing missing samples and introduce a novel tree architecture based on the Long Short-Term Memory (LSTM) networks. In our architecture, we employ a variable number of LSTM networks, which use only the existing inputs in the sequence, in a tree-like architecture without any statistical assumptions or imputations on the missing data. In particular, we incorporate the missingness information by selecting a subset of these LSTM networks based on presence-pattern of a certain number of previous inputs.
LSTM özkodlayıcılar ile ağ yükü tabanlı ihlal tespiti
The increase in the use of computer networks by vast numbers of different devices have allowed malicious entities to develop a plethora of diverse attacks, targeting individuals and businesses. The defence systems need to be kept up to date constantly since new attacks emerge daily, in addition to having a wide range of characteristics. Intrusion detection is a branch of cyber-security that aims to prevent these attacks. Machine learning and deep learning approaches gained popularity in this discipline, as they did in many others such as fraud detection and medicine. Given that network traffic usually displays normal behavior, anomaly detection methods can pinpoint threats by identifying connections with abnormal properties. This task can be accomplished in a supervised or an unsupervised manner. Regardless of the path, constructing meaningful representations of network data is essential. In this thesis, we employ different types of feature extraction methods for computer network data and anomaly detection strategies that can detect malicious behaviour. For the feature extraction task, we aim to obtain vector representations of network payloads such that the core information is more reachable and irrelevant information is discarded. In our setting, the input size can vary due to the nature of the computer network data. Considering this, we use feature extraction methods that can map inputs of varying sizes into feature spaces with fixed dimensionality so that some machine learning approaches, that are otherwise unusable in these settings, can be employed. For the anomaly detection task, we utilize both supervised and unsupervised approaches. The supervised methods make use of the aforementioned feature extraction strategies and use the reduced and fixed dimensional representations of the computer network data. For the unsupervised case, we employ autoencoders that can extract information from sequential data. Recurrent neural networks(RNNs) can process sequential data with varying length. We specifi cally use autoencoders with long short-term memory(LSTM), which is a special form of RNNs with a more complex structure that allows them to handle long-term dependencies in sequential data. Then, anomaly detection is performed using reconstruction error. We conduct experiments using dynamic and realistic data sets, which consist of various types of attacks. Then, we evaluate the validity of our proposed approaches based on AUC and F1 measures.
Yineliyici sinir ağları için verimli çevrimiçi eğitim algoritmaları
Recurrent Neural Networks (RNNs) are widely used for online regression due to their ability to learn nonlinear temporal dependencies. As an RNN model, Long-Short-Term-Memory Networks (LSTMs) are commonly preferred in practice, since these networks are capable of learning long-term dependencies while avoiding the exploding gradient problem. On the other hand, the performance improvement of LSTMs usually comes with the price of their large parameter size, which makes their training significantly demanding in terms of computational and data requirements. In this thesis, we address the computational challenges of LSTM training. We introduce two training algorithms, designed for obtaining the online regression performance of LSTMs with less computational requirements than the state-of-the-art. The introduced algorithms are truly online, i.e., they do not assume any underlying data generating process and future information, except that the dataset is bounded. We discuss theoretical guarantees of the introduced algorithms, along with their asymptotic convergence behavior. Finally, we demonstrate their performance through extensive numerical studies on real and synthetic datasets, and show that they achieve the regression performance of LSTMs with significantly shorter training times.
Kendini düzenleyen noktasal süreç modelleri ile uzay-zamansal dizilerin durağan olmayan dinamiklerini modelleme
We investigate the challenging problem of modeling the non-stationary dynamics of spatio-temporal sequences for prediction applications. Spatio-temporal sequence modeling has critical real-life applications such as natural disaster, social, and criminal event prediction. Even though this problem has been thoroughly studied, many approaches do not address the non-stationarity and sparsity of the spatio-temporal sequences, which are frequently observed in real-life sequences. Here, we introduce a novel prediction algorithm that is capable of modeling non-stationarity in both time and space. Moreover, our algorithm can model both densely and sparsely populated sequences. We partition the spatial region with a decision tree, where each node of the tree corresponds to a subregion. We model the event occurrences in different subregions in space with individual but inter-acting point processes. Our algorithm can jointly optimize the partitioning tree and the interacting point processes through a gradient-based optimization. We compare our approach with statistical models, probabilistic approaches, and deep learning-based approaches, and show that our model achieves the best forecasting performance on real-life datasets such as earthquake and criminal event records.
Derin metrik öğrenmesi ile baştan sona optimize edilebilen gözetimsiz anomali tespiti
We investigate unsupervised anomaly detection for high-dimensional data and introduce a deep metric learning (DML) based framework. In particular, we learn a distance metric through a deep neural network. Through this metric, we project the data into the metric space that better separates the anomalies from the normal data and reduces the effect of the curse of dimensionality for high-dimensional data. We present a novel data distillation method through self-supervision to remedy the conventional practice of assuming all data as normal. We also employ the hard mining technique from the DML literature. We show these components improve the performance of our model. Through an extensive set of experiments on the 14 real-world datasets, our method demonstrates significant performance gains compared to the state-of-the-art unsupervised anomaly detection methods, e.g., an absolute improvement between 4.44% and 11.74% on the average over the 14 datasets. Furthermore, we share the source code of our method on Github to facilitate further research.
Markov anahtarlamalı tekrarlayan yapay sinir ağları ile durağan olmayan zaman serisi tahmini
We investigate nonlinear prediction for nonstationary time series. In most real-life scenarios such as finance, retail, energy and economy applications, time series data exhibits nonstationarity due to the temporally varying dynamics of the underlying system. This situation makes the time series prediction challenging in nonstationary environments. We introduce a novel recurrent neural network (RNN) architecture, which adaptively switches between internal regimes in a Markovian way to model the nonstationary nature of the given data. Our model, Markovian RNN employs a hidden Markov model (HMM) for regime transitions, where each regime controls hidden state transitions of the recurrent cell independently. We jointly optimize the whole network in an end-to-end fashion. We demonstrate the significant performance gains compared to conventional methods such as Markov Switching ARIMA, RNN variants and recent statistical and deep learning-based methods through an extensive set of experiments with synthetic and real-life datasets. We also interpret the inferred parameters and regime belief values to analyze the underlying dynamics of the given sequences.
Zamansal evrişimli sinirsel ağlar için özgün bir geçit mekanizması
We investigate the sequential modeling problem and introduce a novel gating mechanism into the temporal convolutional network architectures. In particular, we propose the Gated Temporal Convolutional Network architecture with elaborately tailored gating mechanisms. In our implementation, we alter the way in which the gradients flow and avoid the vanishing or exploding gradient and the dead ReLU problems. The proposed GTCN architecture is able to model the irregularly sampled sequences as well. In our experiments, we show that the basic GTCN architecture is superior to the generic TCN architectures in various benchmark tasks requiring the modeling of long-term dependencies and irregular sampling intervals. Moreover, we achieve the state-of-the-art results on the permuted sequential MNIST and the sequential CIFAR10 benchmarks with the basic structure.
Uzay-zamansal serilerde analiz ve tahminleme: Hava durumu ve suç tahmininde yeni derin öğrenme mimarileri
We investigate spatiotemporal series through weather and crime forecasting and introduce new successful deep learning architectures that are also applicable to other spatiotemporal domains. First, we present our weather model being an alternative to physical models by addressing the computational cost and the performance. In our weather model, we extend the encoder-decoder structure with the Convolutional Long-short Term Memory units and enhance the performance with attention and context matcher mechanisms. We perform experiments on high-scale, real-life, benchmark numerical weather datasets. We show that attention matrices model atmospheric circulations and successfully capture spatial and temporal relations. Our model obtains the best scores among the baseline deep learning models and comparable results with the physical models. Secondly, we study high-resolution crime prediction and introduce a new generative model with Graph Convolutional Gated Recurrent Units (Graph-ConvGRU) and multivariate Gaussian distributions. We introduce a subdivision algorithm and create a graph representation to tackle the sparsity and complexity problem in high-resolution spatiotemporal data. By leveraging the flexible structure of graph representation, we encode the relations to state vectors for each region. We then create a multivariate probability distribution from the state vectors and perform prediction at any resolution for the first time in the literature. Our model obtains the best score in our experiments on real-life and synthetic datasets compared to the state-of-the-art models. Hence our model is not only generative but also precise. We also provide the source code of our algorithms for reproducibility.
Yeni tecrübe tekrarı mekanizmalarıyla derin deterministik politika gradyanı algoritmalarının performansını artırma
The experience replay mechanism allows agents to use the experiences multiple times. In prior works, the sampling probability of the transitions was adjusted according to their importance. Reassigning sampling probabilities for every transition in the replay buffer after each iteration is highly inefficient. Therefore, experience replay prioritization algorithms recalculate the significance of a transition when the corresponding transition is sampled to gain computational efficiency. However, the importance level of the transitions changes dynamically as the policy and the value function of the agent are updated. In addition, experience replay stores the transitions generated by the previous policies of the agent that may significantly deviate from the most recent policy of the agent. Higher deviation from the most recent policy of the agent leads to more off-policy updates, which is detrimental for the agent. In this thesis, we develop a novel algorithm, Batch Prioritizing Experience Replay via KL Divergence (KLPER), which prioritizes a batch of transitions rather than directly prioritizing each transition. Moreover, to reduce the off-policiness of the updates, our algorithm selects one batch among a certain number of batches and forces the agent to learn through the batch that is most likely generated by the most recent policy of the agent. Also, previous experience replay algorithms in the literature provide the same batches of transitions to the Actor and the Critic Networks of the Deep Deterministic Policy Gradients algorithms. However, the learning principles of these two cascaded components of a deep deterministic policy gradient algorithm contain dissimilarities in terms of their parameter updating strategies. Due to this fact, we attempt to decouple the training of the Actor and the Critic of the deep deterministic policy gradient algorithms in terms of the batches of transitions that they use during the training of the networks. We develop a novel algorithm, Decoupled Prioritized Experience Replay, DPER, that enables the agent to use independently sampled batches of transition for the Actor and the Critic of the Deep Deterministic Policy Gradient Algorithms. DPER utilizes Prioritized Experience Replay, PER, and Batch Prioritizing Experience Replay via KL Divergence, KLPER, to decouple the learning processes of the Critic and the Actor, respectively. We combine our algorithms, KLPER and DPER, with the current state-of-the-art Deep Deterministic Policy Gradient algorithm, DDPG, and TD3, and evaluate it on continuous control tasks. KLPER provides promising improvements for deep deterministic continuous control algorithms in terms of sample efficiency, final performance, and stability of the policy during the training. Moreover, DPER outperforms PER, KLPER, and Vanilla Experience Replay on most of the continuous control tasks. DPER outperforms conventional experience replay strategies without adding a significant amount of computational complexity.
Etkili ardışık veri tahmini için uçtan uca melez mimariler
We investigate nonlinear prediction in an online setting and introduce two hybrid models that effectively mitigate, via end-to-end architectures, the need for hand-designed features and manual model selection issues of conventional nonlinear prediction/regression methods. Particularly, we first use an enhanced recurrent neural network (LSTM) to extract features from sequential signals, while preserving the state information, i.e., the history, and soft gradient boosted decision trees (sGBDT) to produce the final output. The connection is in an end-to-end fashion and we jointly optimize the whole architecture using stochastic gradient descent. Secondly, we again use recursive structures (LSTM) for automatic feature extraction out of raw data but accompany it with a traditional linear time series model (SARIMAX) to deal with the intricacies of the sequential data, e.g., seasonality. The unification of the models is again in a joint manner; it is through a single state space and we optimize the entire architecture using particle filtering. The proposed frameworks are generic so that one can use other recurrent architectures, e.g., GRUs, and differentiable machine learning algorithms as well as time series models that have state space representations in lieu of the specific models presented. We demonstrate the learning behavior of the models on synthetic data and the significant performance improvements over the conventional methods and the disjoint counterparts over various real life datasets, with which we also show the generic nature of the frameworks. Furthermore, we openly share the source code of the proposed methods to facilitate further research.