Machine Learning and Knowledge Discovery in Databases

We present SIDE, a tool for Subjective and Interactive Visual Data Exploration, which lets users explore high dimensional data via subjectively informative 2D data visualizations. Many existing visual analytics tools are either restricted to specific problems and domains or they aim to find visualizations that align with user’s belief about the data. In contrast, our generic tool computes data visualizations that are surprising given a user’s current understanding of the data. The user’s belief state is represented as a set of projection tiles. Hence, this user-awareness offers users an efficient way to interactively explore yet-unknown features of complex high dimensional datasets.

We present GMMbuilder, a tool that allows domain scientists to build Gaussian Mixture Models (GMM) that adhere to domain specific constraints like spatial coherence. Domain experts use this tool to generate different models, extract stable object communities across these models, and use these communities to interactively design a final clustering model that explains the data but also considers prior beliefs and expectations of the domain experts.

Large amounts of multivariate time series data are being generated every day. Understanding this data and finding patterns in it is a contemporary task. To find prominent patterns present in multivariate time series, one can use biclustering, that is looking for patterns both in subsets of variables that show coherent behavior and in a number of time periods. For this, an experimental tool is needed.Here, we present Bipeline, a web-based visualization tool that provides both experts and non-experts with a pipeline for experimenting with multivariate time series biclustering. With Bipeline, it is straightforward to save experiments and try different biclustering algorithms, enabling users to intuitively go from pre-processing to visual analysis of biclusters.

From a molecule to the brain perception, olfaction is a complex phenomenon that remains to be fully understood in neuroscience. Latest studies reveal that the physico-chemical properties of volatile molecules can partly explain the odor perception. Neuroscientists are then looking for new hypotheses to guide their research: physico-chemical descriptors distinguishing a subset of perceived odors. To answer this problem, we present the platform h(odor) that implements descriptive rule discovery algorithms suited for this task. Most importantly, the olfaction experts can interact with the discovery algorithm to guide the search in a huge description space w.r.t their non-formalized background knowledge thanks to an ergonomic user interface.

In this demo we present INSIGHT, a system that provides traffic event detection in Dublin by exploiting Big Data and Crowdsourcing techniques. Our system is able to process and analyze input from multiple heterogeneous urban data sources.

Current technologies allow movements of the players and the ball in football matches to be tracked and recorded with high accuracy and temporal frequency. We demonstrate an approach to analyzing football data with the aim to find typical patterns of spatial arrangement of the field players. It involves transformation of original coordinates to relative positions of the players and the ball with respect to the center and attack vector of each team. From these relative positions, we derive features for characterizing spatial configurations in different time steps during a football game. We apply clustering to these features, which groups the spatial configurations by similarity. By summarizing groups of similar configurations, we obtain representation of spatial arrangement patterns practiced by each team. The patterns are represented visually by density maps built in the teams’ relative coordinate systems. Using additional displays, we can investigate under what conditions each pattern was applied.

By applying spatio-temporal aggregation to traffic data consisting of vehicle trajectories, we generate a spatially abstracted transportation network, which is a directed graph where nodes stand for territory compartments (areas in geographic space) and links (edges) are abstractions of the possible paths between neighboring areas. From time series of traffic characteristics obtained for the links, we reconstruct mathematical models of the interdependencies between the traffic intensity (a.k.a. traffic flow or flux) and mean velocity. Graphical representations of these interdependencies have the same shape as the fundamental diagram of traffic flow through a physical street segment, which is known in transportation science. This key finding substantiates our approach to traffic analysis, forecasting, and simulation leveraging spatial abstraction. We present the process of data-driven generation of traffic forecasting and simulation models, in which each step is supported by visual analytics techniques.

SPMF is an open-source data mining library, specialized in pattern mining, offering implementations of more than 120 data mining algorithms. It has been used in more than 310 research papers to solve applied problems in a wide range of domains from authorship attribution to restaurant recommendation. Its implementations are also commonly used as benchmarks in research papers, and it has also been integrated in several data analysis software programs. After three years of development, this paper introduces the second major revision of the library, named SPMF 2, which provides (1) more than 60 new algorithm implementations (including novel algorithms for sequence prediction), (2) an improved user interface with pattern visualization (3) a novel plug-in system, (4) improved performance, and (5) support for text mining.

We propose a new generator for dynamic attributed networks with community structure which follow the known properties of real-world networks such as preferential attachment, small world and homophily. After the generation, the different graphs forming the dynamic network as well as its evolution can be displayed in the interface. Several measures are also computed to evaluate the properties verified by each graph. Finally, the generated dynamic network, the parameters and the measures can be saved as a collection of files.

The Topy system automates real-time story tracking by utilizing crowd-sourced tagging on social media platforms. Topy employs a state-of-the-art Twitter hashtag recommender to continuously annotate news articles with hashtags, a rich meta-data source that allows connecting articles under drastically different timelines than typical keyword based story tracking systems. Employing social tags for story tracking has the following advantages: (1) social annotation of news enables the detection of emerging concepts and topic drift in a story; (2) hashtags go beyond topics by grouping articles based on connected themes (e.g., #rip, #blacklivesmatter, #icantbreath); (3) hashtags link articles that focus on subplots of the same story (e.g., #palmyra, #isis, #refugeecrisis).

The academic world utterly relies on the concept of scientific collaboration. As in every collaborative network, however, the production of research articles follows hidden co-authoring principles as well as temporal dynamics which generate latent and complex collaboration patterns. In this paper, we present an online advanced tool for real-time rankings of computer scientists under these perspectives.

In this demonstration, we present ReGLL, a system that is able to learn language models taking into account the perceptual context in which the sentences of the model are produced. Thus, ReGLL learns from pairs (Context, Sentence) where: Context is given in the form of an image whose objects have been identified, and Sentence gives a (partial) description of the image. ReGLL uses Inductive Logic Programming Techniques and learns some mappings between n-grams and first order representations of their meanings. The demonstration shows some applications of the language models learned, such as generating relevant sentences describing new images given by the user and translating some sentences from one language to another without the need of any parallel corpus.

We propose a visual analytics tool to support analytic journalists in the exploration of large text corpora. Our tool combines graph modularity-based diagonal biclustering to extract high-level topics with overlapping bi-clustering to elicit fine-grained topic variants. A hybrid topic treemap visualization gives the analyst an overview of all topics. Coordinated sunburst and heatmap visualizations let the analyst inspect and compare topic variants and access document content on demand.

This paper presents a mining system for extracting patterns from Satellite Image Time Series. This system is a fully-fledged tool comprising four main modules for pre-processing, pattern extraction, pattern ranking and pattern visualization. It is based on the extraction of grouped frequent sequential patterns and on swap randomization.

An information retrieval framework is proposed which searches for incident-related social media messages in an automated fashion. Using P2000 messages as an input for this framework and by extracting location information from text, using simple natural language processing techniques, a search for incident-related messages is conducted. A machine learned ranker is trained to create an ordering of the retrieved messages, based on their relevance. This provides an easy accessible interface for emergency response managers to aid them in their decision making process.

We present TwitterCracy, an exploratory search system that allows users to search and monitor across the Twitter streams of political entities. Its exploratory capabilities stem from the application of lightweight time-series based clustering together with biased PageRank to extract facets from tweets and presenting them in a manner that facilitates exploration.

STEM (Science, Technology, Engineering, and Mathematics) fields have become increasingly central to U.S. economic competitiveness and growth. The shortage in the STEM workforce has brought promoting STEM education upfront. The rapid growth of social media usage provides a unique opportunity to predict users’ real-life identities and interests from online texts and photos. In this paper, we propose an innovative approach by leveraging social media to promote STEM education: matching Twitter college student users with diverse LinkedIn STEM professionals using a ranking algorithm based on the similarities of their demographics and interests. We share the belief that increasing STEM presence in the form of introducing career role models who share similar interests and demographics will inspire students to develop interests in STEM related fields and emulate their models. Our evaluation on 2,000 real college students demonstrated the accuracy of our ranking algorithm. We also design a novel implementation that recommends matched role models to the students.

Learning from data streams is a challenge faced by data science professionals from multiple industries. Most of them struggle hardly on applying traditional Machine Learning algorithms to solve these problems. It happens so due to their high availability on ready-to-use software libraries on big data technologies (e.g. SparkML). Nevertheless, most of them cannot cope with the key characteristics of this type of data such as high arrival rate and/or non-stationary distributions. In this paper, we introduce a generic and yet simplistic framework to fill this gap denominated Concept Neurons. It leverages on a combination of continuous inspection schemas and residual-based updates over the model parameters and/or the model output. Such framework can empower the resistance of most of induction learning algorithms to concept drifts. Two distinct and hence closely related flavors are introduced to handle different drift types. Experimental results on successful distinct applications on different domains along transportation industry are presented to uncover the hidden potential of this methodology.

The fast pace of urbanization has given rise to complex transportation networks, such as subway systems, that deploy smart card readers generating detailed transactions of mobility. Predictions of human movement based on these transaction streams represents tremendous new opportunities from optimizing fleet allocation of on-demand transportation such as UBER and LYFT to dynamic pricing of services. However, transportation research thus far has primarily focused on tackling other challenges from traffic congestion to network capacity. To take on this new opportunity, we propose a real-time framework, called PULSE (Prediction Framework For Usage Load on Subway SystEms), that offers accurate multi-granular arrival crowd flow prediction at subway stations. PULSE extracts and employs two types of features such as streaming features and station profile features. Streaming features are time-variant features including time, weather, and historical traffic at subway stations (as time-series of arrival/departure streams), where station profile features capture the time-invariant unique characteristics of stations, including each station’s peak hour crowd flow, remoteness from the downtown area, and mean flow. Then, given a future prediction interval, we design novel stream feature selection and model selection algorithms to select the most appropriate machine learning models for each target station and tune that model by choosing an optimal subset of stream traffic features from other stations. We evaluate our PULSE framework using real transaction data of 11 million passengers from a subway system in Shenzhen, China. The results demonstrate that PULSE greatly improves the accuracy of predictions at all subway stations by up to $$49\,\%$$ over baseline algorithms.

Co-evolving patterns exist in many Spatial-temporal time series Data, which shows invaluable information about evolving patterns of the data. However, due to the sensor readings’ spatial and temporal heterogeneity, how to find the stable and dynamic co-evolving zones remains an unsolved issue. In this paper, we proposed a novel divide-and-conquer strategy to find the dynamic co-evolving zones that systematically leverages the heterogeneity challenges. The precision of spatial inference and temporal prediction improved by 7 % and 8 % respectively by using the found patterns, which shows the effectiveness of the found patterns. The system has also been deployed with the Haidian Ministry of Environmental Protection, Beijing, China, providing accurate spatial-temporal predictions and help the government make more scientific strategies for environment treatment.

Because of user movements and activities, heartbeats recorded from wearable devices typically feature a large degree of variability in their morphology. Learning problems, which in ECG monitoring often involve learning a user-specific model to describe the heartbeat morphology, become more challenging.Our study, conducted on ECG tracings acquired from the Pulse Sensor – a wearable device from our industrial partner – shows that dictionaries yielding sparse representations can successfully model heartbeats acquired in typical wearable-device settings. In particular, we show that sparse representations allow to effectively detect heartbeats having an anomalous morphology. Remarkably, the whole ECG monitoring can be executed online on the device, and the dictionary can be conveniently reconfigured at each device positioning, possibly relying on an external host.

Local businesses and retail stores are a crucial part of local economy. Local governments design policies for facilitating the growth of these businesses that can consequently have positive externalities on the local community. However, many times these policies have completely opposite from the expected results (e.g., free curb parking instead of helping businesses has been illustrated to actually hurt them due to the small turnover per spot). Hence, it is important to evaluate the outcome of such policies in order to provide educated decisions for the future. In the era of social and ubiquitous computing, mobile social media, such as Foursquare, form a platform that can help towards this goal. Data from these platforms capture semantic information of human mobility from which we can distill the potential economic activities taking place. In this paper we focus on street fairs (e.g., arts festivals) and evaluate their ability to boost economic activities in their vicinity. In particular, we collected data from Foursquare for the three month period between June 2015 and August 2015 from the city of Pittsburgh. During this period several street fairs took place. Using these events as our case study we analyzed the data utilizing propensity score matching and a quasi-experimental technique inspired by the difference-in-differences method. Our results indicate that street fairs provide positive externalities to nearby businesses. We further analyzed the spatial reach of this impact and we find that it can extend up to 0.6 miles from the epicenter of the event.

Urban data management is already an essential element of modern cities. The authorities can build on the variety of automatically generated information and develop intelligent services that improve citizens daily life, save environmental resources or aid in coping with emergencies. From a data mining perspective, urban data introduce a lot of challenges. Data volume, velocity and veracity are some obvious obstacles. However, there are even more issues of equal importance like data quality, resilience, privacy and security. In this paper we describe the development of a set of techniques and frameworks that aim at effective and efficient urban data management in real settings. To do this, we collaborated with the city of Dublin and worked on real problems and data. Our solutions were integrated in a system that was evaluated and is currently utilized by the city.

Cardiac computed tomography is a non-invasive technique to image the beating heart. One of the main concerns during the procedure is the total radiation dose imposed on the patient. Prospective electrocardiographic (ECG) gating methods may notably reduce the radiation exposure. However, very few investigations address accompanying problems encountered in practice. Several types of unique non-biological factors, such as the dynamic electrical field induced by rotating components in the scanner, influence the ECG and can result in artifacts that can ultimately cause prospective ECG gating algorithms to fail. In this paper, we present an approach to automatically detect non-biological artifacts within ECG signals, acquired in this context. Our solution adapts discord discovery, robust PCA, and signal processing methods for detecting such disturbances. It achieved an average area under the precision-recall curve (AUPRC) and receiver operating characteristics curve (AUROC) of 0.996 and 0.997 in our cross-validation experiments based on 2,581 ECGs. External validation on a separate hold-out dataset of 150 ECGs, annotated by two domain experts (88 % inter-expert agreement), yielded average AUPRC and AUROC scores of 0.890 and 0.920. Our solution is deployed to automatically detect non-biological anomalies within a continuously updated database, currently holding over 120,000 ECGs.

A major focus of the commercial aviation community is discovery of unknown safety events in flight operations data. Data-driven unsupervised anomaly detection methods are better at capturing unknown safety events compared to rule-based methods which only look for known violations. However, not all statistical anomalies that are discovered by these unsupervised anomaly detection methods are operationally significant (e.g., represent a safety concern). Subject Matter Experts (SMEs) have to spend significant time reviewing these statistical anomalies individually to identify a few operationally significant ones. In this paper we propose an active learning algorithm that incorporates SME feedback in the form of rationales to build a classifier that can distinguish between uninteresting and operationally significant anomalies. Experimental evaluation on real aviation data shows that our approach improves detection of operationally significant events by as much as 75 % compared to the state-of-the-art. The learnt classifier also generalizes well to additional validation data sets.

We address the problem of detecting whether an engine is misfiring by using machine learning techniques on transformed audio data collected from a smartphone. We recorded audio samples in an uncontrolled environment and extracted Fourier, Wavelet and Mel-frequency Cepstrum features from normal and abnormal engines. We then implemented Fisher Score and Relief Score based variable ranking to obtain an informative reduced feature set for training and testing classification algorithms. Using this feature set, we were able to obtain a model accuracy of over 99 % using a linear SVM applied to outsample data. This application of machine learning to vehicle subsystem monitoring simplifies traditional engine diagnostics, aiding vehicle owners in the maintenance process and opening up new avenues for pervasive mobile sensing and automotive diagnostics.

The amount of textual data available from digitalized sources such as free online libraries or social media posts has increased drastically in the last decade. In this paper, the main idea to analyze authors by their grammatical writing style is presented. In particular, tasks like authorship attribution, plagiarism detection or author profiling are tackled using the presented algorithm, revealing promising results. Thereby all of the presented approaches are ultimately solved by machine learning algorithms.

The acceptance of analytical methods for discrimination discovery by practitioners and legal scholars can be only achieved if the data mining and machine learning communities will be able to provide case studies, methodological refinements, and the consolidation of a KDD process. We summarize here an approach along these directions.

Applications are getting increasingly interconnected. Al-though the interconnectedness provide new ways to gather information about the user, not all user information is ready to be directly implemented in order to provide a personalized experience to the user. Therefore, a general model is needed to which users’ behavior, preferences, and needs can be connected to. In this paper we present our works on a personality-based music recommender system in which we use users’ personality traits as a general model. We identified relationships between users’ personality and their behavior, preferences, and needs, and also investigated different ways to infer users’ personality traits from user-generated data of social networking sites (i.e., Facebook, Twitter, and Instagram). Our work contributes to new ways to mine and infer personality-based user models, and show how these models can be implemented in a music recommender system to positively contribute to the user experience.

Recurrence quantification analysis (RQA) was developed in order to quantify differently appearing recurrence plots (RPs) based on their small-scale structures, which generally indicate the number and duration of recurrences in a dynamical system. Although RQA measures are traditionally employed in analyzing complex systems and identifying transitions, recent work has shown that they can also be used for pairwise dissimilarity comparisons of time series. We explain why RQA is not only a modern method for nonlinear data analysis but also is a very promising technique for various time series mining tasks.

Today’s steel industry is characterized by overcapacity and increasing competitive pressure. There is a need for continuously improving processes, with a focus on consistent enhancement of efficiency, improvement of quality and thereby better competitiveness. About 70 % of steel is produced using the BF-BOF (Blast Furnace - Blow Oxygen Furnace) route worldwide. The BOF is the first step of controlling the composition of the steel and has an impact on all further processing steps and the overall quality of the end product. Multiple sources of process-related variance and overall harsh conditions for sensors and automation systems in general lead to a process complexity that is not easy to model with thermodynamic or metallurgical approaches. In this paper we want to give an insight how to improve the output quality with machine learning based modeling and which constraints and requirements are necessary for an online application in real-time.

In software project planning project managers have to keep track of several things simultaneously including the estimation of the consequences of decisions about, e.g., the team constellation. The application of machine learning techniques to predict possible outcomes is a widespread research topic in software engineering. In this paper, we summarize our work in the field of learning from project history.

Inverse reinforcement learning (irl) provides a concise framework for learning behaviors from human demonstrations; and is highly desired in practical and difficult to specify tasks such as normative robot navigation. However, most existing irl algorithms are often ladened with practical challenges such as representation mismatch and poor scalability when deployed in real world tasks. Moreover, standard reinforcement learning (rl) representations often do not allow for incorporation of task constraints common for example in robot navigation. In this paper, we present an approach that tackles these challenges in a unified manner and delivers a learning setup that is both practical and scalable. We develop a graph-based spare representation for rl and a scalable irl algorithm based on sampled trajectories. Experimental evaluation in simulation and from a real deployment in a busy airport demonstrate the strengths of the learning setup over existing approaches.

Recent technological advances in the fields of biology and medicine allow measuring single cells into unprecedented depth. This results in new types of high-throughput datasets that shed new lights on cell development, both in healthy as well as diseased tissues. However, studying these biological processes into greater detail crucially depends on novel computational techniques that efficiently mine single cell data sets. In this paper, we introduce machine learning techniques for single cell data analysis: we summarize the main developments in the field, and highlight a number of interesting new avenues that will likely stimulate the design of new types of machine learning algorithms.

Most classification algorithms are aimed at predicting the value or values of a single target (class) attribute. However, some real-world classification tasks involve several targets that need to be predicted simultaneously. The Multi-objective Info-Fuzzy Network (M-IFN) algorithm builds an ordered (oblivious) decision-tree model for a multi-target classification task. After summarizing the principles and the properties of the M-IFN algorithm, this paper reviews three case studies of applying M-IFN to practical problems in industry and science.

Search behavior, and information seeking behavior more generally, is often motivated by tasks that prompt search processes that are often lengthy, iterative, and intermittent, and are characterized by distinct stages, shifting goals and multitasking. Current search systems do not provide adequate support for users tackling complex tasks due to which the cognitive burden of keeping track of such tasks is placed on the searcher. In this note, we summarize our recent efforts towards extracting search tasks from search logs. Based on recent advancements in Bayesian Nonparametrics and distributional semantics, we propose novel algorithms to extract task and subtasks from a query collection. The models discussed can inform the design of the next generation of task-based search systems that leverage user’s task behavior for better support and personalization.

As a popular data mining tasks, frequent pattern mining discovers implicit, previously unknown and potentially useful knowledge in the form of sets of frequently co-occurring items or events. Many existing data mining algorithms return to users with long textual lists of frequent patterns, which may not be easily comprehensible. As a picture is worth a thousand words, having a visual means for humans to interact with computers would be beneficial. This is when human-computer interaction (HCI) research meets data mining research. In particular, the popular HCI task of data and result visualization could help data miners to visualize the original data and to analyze the mined results (in the form of frequent patterns). In this paper, we present a few systems for data and visual analytics of frequent patterns, which integrate (i) data analytics and mining with (ii) data and result visualization.

Recent years have seen a significant increase in the number of applications requiring accurate and up-to-date spatial data. In this context crowdsourced maps such as OpenStreetMap (OSM) have the potential to provide a free and timely representation of our world. However, one factor that negatively influences the proliferation of these maps is the uncertainty about their data quality. This paper presents structured and unstructured machine learning methods to automatically assess and improve the semantic quality of streets in the OSM database.

Local exceptionality detection on social interaction networks includes the analysis of resources created by humans (e. g., social media) as well as those generated by sensor devices in the context of (complex) interactions. This paper provides a structured overview on a line of work comprising a set of papers that focus on data-driven exploration and modeling in the context of social network analysis, community detection and pattern mining.