We study active correlation clustering where pairwise similarities are not provided upfront and must be queried in a cost-efficient manner through active learning. Specifically, we focus on the cold-start scenario, where no true initial pairwise similarities are available for active learning. To address this challenge, we propose a coverage-aware method that encourages diversity early in the process. We demonstrate the effectiveness of our approach through several synthetic and real-world experiments.

Bottleneck identification is a challenging task in network analysis, especially when the network is not fully specified. To address this task, we develop a unified online learning framework based on combinatorial semi-bandits that performs bottleneck identification alongside learning the specifications of the underlying network. Within this framework, we adapt and investigate several combinatorial semi-bandit methods such as epsilon-greedy, LinUCB, BayesUCB, and Thompson Sampling. Our framework is able to employ contextual information in the form of contextual bandits. We evaluate our framework on the real-world application of road networks and demonstrate its effectiveness in different settings.

A well-studied problem in data mining is large itemset mining. To address this problem over very large datasets, several approximate algorithms have been introduced, where an important class of such methods relies on sampling. However in the literature, only methods that are based on uniform sampling are investigated. In this paper, first we discuss how different sampling methods can be described using a generic sampling algorithm and study a property desirable for sampling methods. Then we use this property to argue that some non-uniform sampling methods may work better. We accordingly propose methods that sample each transaction proportional to its number of items or proportional to its number of frequent items. Finally, by conducting extensive experiments over real-world datasets, we show that non-uniform sampling methods usually outperform the uniform method.

Determinantal point processes (DPPs) have become prominent in data summarization and recommender system tasks for their ability to simultaneously model diversity as well as relevance. In practical applications, k-Determinantal point processes (k-DPPs) are used to yield a selection of k items from a set of size N that are the most representative of the set. In this paper, we study a special case of the diverse subset selection problem where a fixed set GO is already given as a forced recommendation and the task is to determine the remainder of the recommendation G1. The standard k-DPP optimization objectives here can suggest items that are close to optimal when considering only items in G1, but are arbitrarily close to items in G0, i.e., they might not be sufficiently diverse w.r.t. G0. We explore a semi-supervised k-DPP objective that simultaneously considers G0 and G1 and compares the difference between the two recommendations. We demonstrate our findings using multiple examples where the diverse subset selection problem with forced recommendation is important in practice.

We employ Reinforcement Learning (RL) techniques with improved state and action spaces for tactical decision making in an autonomous truck. Specifically, we implement Adaptive Cruise Control (ACC) and lane change maneuvers for the autonomous truck in a highway scenario. We show the results obtained using three reinforcement learning algorithms: Deep Q-Network (DQN), Advantage Actor-Critic (A2C) and Proximal Policy Optimization (PPO). Our results demonstrate that it is beneficial to separate high-level decision-making processes and low-level control actions between the RL agent and the low-level controllers based on physical models. Furthermore, we design a realistic reward function based on the Total Cost of Operation (TCOP) of the truck to guide the RL agent towards optimal driving strategy.

Acquiring labelled data for machine learning tasks, for example, for software performance prediction, remains a resource-intensive task. This study extends our previous work by introducing a batch-mode deep active learning approach tailored for regression in graph-structured data. Our framework leverages the source code conversion into Flow Augmented-AST graphs (FA-AST), subsequently utilizing both supervised and unsupervised graph embeddings. In contrast to single-instance querying, the batch-mode paradigm adaptively selects clusters of unlabeled data for labelling. We deploy an array of base kernels, kernel transformations, and selection methods, informed by both Bayesian and non-Bayesian strategies, to enhance the sample efficiency of neural network regression. Our experimental evaluation, conducted on multiple real-world software performance datasets, demonstrates the efficacy of the batch mode deep active learning approach in achieving robust performance with a reduced labelling budget. The methodology scales effectively to larger datasets and requires minimal alterations to existing neural network architectures.

With the growing rate of online social networks, the number of fake accounts is multiplying day by day. There exist many approaches in the literature that try to distinguish fake accounts from real ones, for example, those that use machine learning and classification techniques to learn whether a user should be labeled as fake (bot) or not. In this paper, we follow a different approach and try to use node measurements in the field of complex networks analysis to identify fake accounts. We first model users’ interactions with a large graph. For example, in Twitter, we can form graphs of follower-following, comments, retweets, mentions, and so on. We then investigate different measurements, such as centrality indices and their correlations, to separate real and fake accounts. We find that measurements such as average path length, eigenvector centrality, harmonic centrality, degree, local reaching centrality and their correlations provide good indicators to distinguish real and fake accounts.

A technique that has recently become popular to analyze graph data is node embedding learning. Many graph problems, such as node classification, link prediction and node clustering, can be solved using these embeddings. However, in the literature the efficiency of different embedding generation algorithms paired with different clustering algorithms is not extensively investigated. In this paper, we study the efficiency of well-known embedding generation algorithms in combination with clustering algorithms, to detect communities. We consider four embedding generation algorithms that use mechanisms such as convolution, attention, inductivity and shallowness; and three popular data clustering algorithms. Our experimental results reveal that the combination of GraphSAGE (that uses inductivity) with KMeans++ yields the best results. This can be due to high quality of embedding vectors generated by GraphSAGE and the regular shapes that the embeddings find in the vector space.

Recent developments in artificial intelligence and automation support a new drug design paradigm: autonomous drug design. Under this paradigm, generative models can provide suggestions on thousands of molecules with specific properties, and automated laboratories can potentially make, test and analyze molecules with minimal human supervision. However, since still only a limited number of molecules can be synthesized and tested, an obvious challenge is how to efficiently select among provided suggestions in a closed-loop system. We formulate this task as a stochastic multi-armed bandit problem with multiple plays, volatile arms and similarity information. To solve this task, we adapt previous work on multi-armed bandits to this setting, and compare our solution with random sampling, greedy selection and decaying-epsilon-greedy selection strategies. According to our simulation results, our approach has the potential to perform better exploration and exploitation of the chemical space for autonomous drug design.

We study Frank-Wolfe algorithms – standard, pairwise, and away-steps – for efficient optimization of Dominant Set Clustering. We present a unified and computationally efficient framework to employ the different variants of Frank-Wolfe methods, and we investigate its effectiveness via several experimental studies. In addition, we provide explicit convergence rates for the algorithms in terms of the so-called Frank-Wolfe gap. The theoretical analysis has been specialized to Dominant Set Clustering and covers consistently the different variants.

Knowledge transfer is shown to be a very successful technique for training neural classifiers: together with the ground truth data, it uses the "privileged information" (PI) obtained by a "teacher" network to train a "student" network. It has been observed that classifiers learn much faster and more reliably via knowledge transfer. However, there has been little or no theoretical analysis of this phenomenon. To bridge this gap, we propose to approach the problem of knowledge transfer by regularizing the fit between the teacher and the student with PI provided by the teacher. Using tools from dynamical systems theory, we show that when the student is an extremely wide two layer network, we can analyze it in the kernel regime and show that it is able to interpolate between PI and the given data. This characterization sheds new light on the relation between the training error and capacity of the student relative to the teacher. Another contribution of the paper is a quantitative statement on the convergence of student network. We prove that the teacher reduces the number of required iterations for a student to learn, and consequently improves the generalization power of the student. We give corresponding experimental analysis that validates the theoretical results and yield additional insights.

We study different aspects of active learning with deep neural networks in a consistent and unified way. i) We investigate incremental and cumulative training modes which specify how the newly labeled data are used for training. ii) We study active learning w.r.t. the model configurations such as the number of epochs and neurons as well as the choice of batch size. iii) We consider in detail the behavior of query strategies and their corresponding informativeness measures and accordingly propose more efficient querying procedures. iv) We perform statistical analyses, e.g., on actively learned classes and test error estimation, that reveal several insights about active learning. v) We investigate how active learning with neural networks can benefit from pseudo-labels as proxies for actual labels.

In recent years, learning embeddings for nodes of a graph has become one of the most efficient w ays t o solve different graph problems such as link prediction, clustering and classification. I n t his p aper, w e p ropose a n ovel m ethod, called SECI, for learning embeddings of nodes, with application to link prediction. SECI samples from the network using breadth-first search and depth-first s earch, a nd i nterpolates b etween these two using centrality indices. The intuition behind SECI is that for nodes that have a low centrality score only a very small neighborhood is explored; and for dominant nodes that have a high centrality score a large neighborhood is explored. We evaluate the empirical performance of SECI over several realworld networks and show that it outperforms well-known existing algorithms.

In order to yield a more balanced partitioning, we investigate the use of additive regularizations for the Min Cut cost function, instead of normalization. In particular, we study the case where the regularization term is the sum of the squared size of the clusters, which then leads to shifting (adaptively) the pairwise similarities. We study the connection of such a model with Correlation Clustering and then propose an efficient local search optimization algorithm to solve the new clustering problem. Finally, we demonstrate the superior performance of our method by extensive experiments on different datasets.

We study efficient computation of Minimax distances measures, which enable to capture the correct structures via taking the transitive relations into account. We analyze in detail two settings, the dense graphs and the sparse graphs. In particular, we show that an adapted variant of the Kruskal's algorithm is the most efficient approach for computing pairwise Minimax distances. However, for dense graphs we require a preprocessing step based on the Prim's algorithm, in order to reduce the set of candidate edges to be investigated. For each case, we study the correctness, efficiency and computational optimality of our approach. We perform numerical experiments on several datasets to validate the superior performance of our methods.

We study Minimax distance measures for K-nearest neighbor search and classification. Recently, the use of this distance measure is shown to improve the K-nearest neighbor classification results. We consider the computational aspects of this problem and propose an efficient and general-purpose algorithm for computing Minimax neighbors which requires a significantly lower runtime and is applicable with any arbitrary distance measure. We study the computational optimality of our approach and its connection to the Prim's algorithm, and then, generalize our analysis to computing one-to-all Minimax distances. In the following, we investigate in detail the edges selected by Minimax distances and thereby explore the ability of Minimax distances in detecting outlier objects. We evaluate the performance of our methods on a variety of real-world datasets, e.g. text documents and images.

Mining frequent tree patterns has many practical applications in areas such as XML document mining, Web mining, bioinformatics, network routing and so on. Most of the previous works used an apriori-based approach for candidate generation and frequency counting in their algorithms. In these approaches the state space grows exponentially since many unreal candidates are generated, especially when there are lots of large patterns among the data. To tackle these problems, we propose TDU, a top-down approach for mining all maximal, labeled, unordered, and embedded subtrees from a collection of tree-structured data. We would evaluate the effectiveness of the TDU algorithm in comparison to the previous works

Tree structures have gained popularity for storing data from different domains such as XML documents, bio informatics and so on. Clustering these data can facilitate different operations. In this paper, we propose TreeCluster, a novel and heuristic algorithm for clustering tree structured data. This algorithm considers a representative tree for each cluster. For each input tree T, TreeCluster computes the composition of the tree T and each of the clusters. Tree T belongs to the cluster which its composed tree gains the best score. After adding a tree to a cluster the representative tree of that cluster is updated. We evaluate the accuracy of the TreeCluster algorithm in comparison to the previous works