Machine Learning and Knowledge Discovery in Databases

With the rapid growth of e-commerce, a large number of online transactions are processed every day. In this paper, we take the initiative to conduct a systematic study of the challenging prediction problems of sales bursts. Here, we propose a novel model to detect bursts, find the bursty features, namely the start time of the burst, the peak value of the burst and the off-burst value, and predict the entire burst shape. Our model analyzes the features of similar sales bursts in the same category, and applies them to generate the prediction. We argue that the framework is capable of capturing the seasonal and categorical features of sales burst. Based on the real data from JD.com, we conduct extensive experiments and discover that the proposed model makes a relative MSE improvement of 71% and 30% over LSTM and ARMA.

Attribution studies in climate science aim for scientifically ascertaining the influence of climatic variations on natural or anthropogenic factors. Many of those studies adopt the concept of Granger causality to infer statistical cause-effect relationships, while utilizing traditional autoregressive models. In this article, we investigate the potential of state-of-the-art time series classification techniques to enhance causal inference in climate science. We conduct a comparative experimental study of different types of algorithms on a large test suite that comprises a unique collection of datasets from the area of climate-vegetation dynamics. The results indicate that specialized time series classification methods are able to improve existing inference procedures. Substantial differences are observed among the methods that were tested.

Visual animal biometrics is rapidly gaining popularity as it enables a non-invasive and cost-effective approach for wildlife monitoring applications. Widespread usage of camera traps has led to large volumes of collected images, making manual processing of visual content hard to manage. In this work, we develop a framework for automatic detection and recognition of individuals in different patterned species like tigers, zebras and jaguars. Most existing systems primarily rely on manual input for localizing the animal, which does not scale well to large datasets. In order to automate the detection process while retaining robustness to blur, partial occlusion, illumination and pose variations, we use the recently proposed Faster-RCNN object detection framework to efficiently detect animals in images. We further extract features from AlexNet of the animal’s flank and train a logistic regression (or Linear SVM) classifier to recognize the individuals. We primarily test and evaluate our framework on a camera trap tiger image dataset that contains images that vary in overall image quality, animal pose, scale and lighting. We also evaluate our recognition system on zebra and jaguar images to show generalization to other patterned species. Our framework gives perfect detection results in camera trapped tiger images and a similar or better individual recognition performance when compared with state-of-the-art recognition techniques.

Accurate electricity load forecasting is of crucial importance for power system operation and smart grid energy management. Different factors, such as weather conditions, lagged values, and day types may affect electricity load consumption. We propose to use multiple kernel learning (MKL) for electricity load forecasting, as it provides more flexibilities than traditional kernel methods. Computation time is an important issue for short-term load forecasting, especially for energy scheduling demand. However, conventional MKL methods usually lead to complicated optimization problems. Another practical aspect of this application is that there may be very few data available to train a reliable forecasting model for a new building, while at the same time we may have prior knowledge learned from other buildings. In this paper, we propose a boosting based framework for MKL regression to deal with the aforementioned issues for short-term load forecasting. In particular, we first adopt boosting to learn an ensemble of multiple kernel regressors, and then extend this framework to the context of transfer learning. Experimental results on residential data sets show the effectiveness of the proposed algorithms.

Clostridium difficile infection (CDI) is a common hospital acquired infection with a $1B annual price tag that resulted in $$\sim $$30,000 deaths in 2011. Studies have shown that early detection of CDI significantly improves the prognosis for the individual patient and reduces the overall mortality rates and associated medical costs. In this paper, we present CREST: CDI Risk Estimation, a data-driven framework for early and continuous detection of CDI in hospitalized patients. CREST uses a three-pronged approach for high accuracy risk prediction. First, CREST builds a rich set of highly predictive features from Electronic Health Records. These features include clinical and non-clinical phenotypes, key biomarkers from the patient’s laboratory tests, synopsis features processed from time series vital signs, and medical history mined from clinical notes. Given the inherent multimodality of clinical data, CREST bins these features into three sets: time-invariant, time-variant, and temporal synopsis features. CREST then learns classifiers for each set of features, evaluating their relative effectiveness. Lastly, CREST employs a second-order meta learning process to ensemble these classifiers for optimized estimation of the risk scores. We evaluate the CREST framework using publicly available critical care data collected for over 12 years from Beth Israel Deaconess Medical Center, Boston. Our results demonstrate that CREST predicts the probability of a patient acquiring CDI with an AUC of 0.76 five days prior to diagnosis. This value increases to 0.80 and even 0.82 for prediction two days and one day prior to diagnosis, respectively.

When will a server fail catastrophically in an industrial datacenter? Is it possible to forecast these failures so preventive actions can be taken to increase the reliability of a datacenter? To answer these questions, we have studied what are probably the largest, publicly available datacenter traces, containing more than 104 million events from 12,500 machines. Among these samples, we observe and categorize three types of machine failures, all of which are catastrophic and may lead to information loss, or even worse, reliability degradation of a datacenter. We further propose a two-stage framework—DC-Prophet (DC-Prophet stands for DataCenter-Prophet.)—based on One-Class Support Vector Machine and Random Forest. DC-Prophet extracts surprising patterns and accurately predicts the next failure of a machine. Experimental results show that DC-Prophet achieves an AUC of 0.93 in predicting the next machine failure, and a F₃-score (The ideal value of F₃-score is 1, indicating perfect predictions. Also, the intuition behind F₃-score is to value “Recall” about three times more than “Precision” [12].) of 0.88 (out of 1). On average, DC-Prophet outperforms other classical machine learning methods by 39.45% in F₃-score.

Successful inventory management in retail entails accurate demand forecasts for many weeks/months ahead. Forecasting models use seasonality: recurring pattern of sales every year, to make this forecast. In e-commerce setting, where the catalog of items is much larger than brick and mortar stores and hence includes a lot of items with short history, it is infeasible to compute seasonality for items individually. It is customary in these cases to use ideas from factor analysis and express seasonality by a few factors/basis vectors computed together for an entire assortment of related items. In this paper, we demonstrate the effectiveness of choosing vectors with disjoint support as basis for seasonality when dealing with a large number of short time-series. We give theoretical results on computation of disjoint support factors that extend the state of the art, and also discuss temporal regularization necessary to make it work on walmart e-commerce dataset. Our experiments demonstrate a marked improvement in forecast accuracy for items with short history.

Identifying events in real-time data streams such as Twitter is crucial for many occupations to make timely, actionable decisions. It is however extremely challenging because of the subtle difference between “events” and trending topics, the definitive rarity of these events, and the complexity of modern Internet’s text data. Existing approaches often utilize topic modeling technique and keywords frequency to detect events on Twitter, which have three main limitations: (1) supervised and semi-supervised methods run the risk of missing important, breaking news events; (2) existing topic/event detection models are base on words, while the correlations among phrases are ignored; (3) many previous methods identify trending topics as events. To address these limitations, we propose the model, PhraseNet, an algorithm to detect and summarize events from tweets. To begin, all topics are defined as a clustering of high-frequency phrases extracted from text. All trending topics are then identified based on temporal spikes of the phrase cluster frequencies. PhraseNet thus filters out high-confidence events from other trending topics using number of peaks and variance of peak intensity. We evaluate PhraseNet on a three month duration of Twitter data and show the both the efficiency and the effectiveness of our approach.

Random forests are among the most popular classification and regression methods used in industrial applications. To be effective, the parameters of random forests must be carefully tuned. This is usually done by choosing values that minimize the prediction error on a held out dataset. We argue that error reduction is only one of several metrics that must be considered when optimizing random forest parameters for commercial applications. We propose a novel metric that captures the stability of random forest predictions, which we argue is key for scenarios that require successive predictions. We motivate the need for multi-criteria optimization by showing that in practical applications, simply choosing the parameters that lead to the lowest error can introduce unnecessary costs and produce predictions that are not stable across independent runs. To optimize this multi-criteria trade-off, we present a new framework that efficiently finds a principled balance between these three considerations using Bayesian optimisation. The pitfalls of optimising forest parameters purely for error reduction are demonstrated using two publicly available real world datasets. We show that our framework leads to parameter settings that are markedly different from the values discovered by error reduction metrics alone.

In traditional A/B testing, we have two variants of the same product, a pool of test subjects, and a measure of success. In a randomized experiment, each test subject is presented with one of the two variants, and the measure of success is aggregated per variant. The variant of the product associated with the most success is retained, while the other variant is discarded. This, however, presumes that the company producing the products only has enough capacity to maintain one of the two product variants. If more capacity is available, then advanced data science techniques can extract more profit for the company from the A/B testing results. Exceptional Model Mining is one such advanced data science technique, which specializes in identifying subgroups that behave differently from the overall population. Using the association model class for EMM, we can find subpopulations that prefer variant A where the general population prefers variant B, and vice versa. This data science technique is applied on data from StudyPortals, a global study choice platform that ran an A/B test on the design of aspects of their website.

Understanding the complex dynamics in the real-world such as in multi-agent behaviors is a challenge in numerous engineering and scientific fields. Spectral analysis using Koopman operators has been attracting attention as a way of obtaining a global modal description of a nonlinear dynamical system, without requiring explicit prior knowledge. However, when applying this to the comparison or classification of complex dynamics, it is necessary to incorporate the Koopman spectra of the dynamics into an appropriate metric. One way of implementing this is to design a kernel that reflects the dynamics via the spectra. In this paper, we introduced Koopman spectral kernels to compare the complex dynamics by generalizing the Binet-Cauchy kernel to nonlinear dynamical systems without specifying an underlying model. We applied this to strategic multiagent sport plays wherein the dynamics can be classified, e.g., by the success or failure of the shot. We mapped the latent dynamic characteristics of multiple attacker-defender distances to the feature space using our kernels and then evaluated the scorability of the play by using the features in different classification models.

Mobile phone metadata is increasingly used for humanitarian purposes in developing countries as traditional data is scarce. Basic demographic information is however often absent from mobile phone datasets, limiting the operational impact of the datasets. For these reasons, there has been a growing interest in predicting demographic information from mobile phone metadata. Previous work focused on creating increasingly advanced features to be modeled with standard machine learning algorithms. We here instead model the raw mobile phone metadata directly using deep learning, exploiting the temporal nature of the patterns in the data. From high-level assumptions we design a data representation and convolutional network architecture for modeling patterns within a week. We then examine three strategies for aggregating patterns across weeks and show that our method reaches state-of-the-art accuracy on both age and gender prediction using only the temporal modality in mobile metadata. We finally validate our method on low activity users and evaluate the modeling assumptions.

E-commerce websites such as Amazon, Alibaba, Flipkart, and Walmart sell billions of products. Machine learning (ML) algorithms involving products are often used to improve the customer experience and increase revenue, e.g., product similarity, recommendation, and price estimation. The products are required to be represented as features before training an ML algorithm. In this paper, we propose an approach called MRNet-Product2Vec for creating generic embeddings of products within an e-commerce ecosystem. We learn a dense and low-dimensional embedding where a diverse set of signals related to a product are explicitly injected into its representation. We train a Discriminative Multi-task Bidirectional Recurrent Neural Network (RNN), where the input is a product title fed through a Bidirectional RNN and at the output, product labels corresponding to fifteen different tasks are predicted. The task set includes several intrinsic characteristics about a product such as price, weight, size, color, popularity, and material. We evaluate the proposed embedding quantitatively and qualitatively. We demonstrate that they are almost as good as sparse and extremely high-dimensional TF-IDF representation in spite of having less than 3% of the TF-IDF dimension. We also use a multimodal autoencoder for comparing products from different language-regions and show preliminary yet promising qualitative results.

The process of liquidity provision in financial markets can result in prolonged exposure to illiquid instruments for market makers. In this case, where a proprietary position is not desired, pro-actively targeting the right client who is likely to be interested can be an effective means to offset this position, rather than relying on commensurate interest arising through natural demand. In this paper, we consider the inference of a client profile for the purpose of corporate bond recommendation, based on typical recorded information available to the market maker. Given a historical record of corporate bond transactions and bond meta-data, we use a topic-modelling analogy to develop a probabilistic technique for compiling a curated list of client recommendations for a particular bond that needs to be traded, ranked by probability of interest. We show that a model based on Latent Dirichlet Allocation offers promising performance to deliver relevant recommendations for sales traders.

Timely identification of dissatisfied customers would provide corporations and other customer serving enterprises the opportunity to take meaningful interventions. This work describes a fully operational system we have developed at a large US insurance company for predicting customer satisfaction following all incoming phone calls at our call center. To capture call relevant information, we integrate signals from multiple heterogeneous data sources including: speech-to-text transcriptions of calls, call metadata (duration, waiting time, etc.), customer profiles and insurance policy information. Because of its ordinal, subjective, and often highly-skewed nature, self-reported survey scores presents several modeling challenges. To deal with these issues we introduce a novel modeling workflow: First, a ranking model is trained on the customer call data fusion. Then, a convolutional fitting function is optimized to map the ranking scores to actual survey satisfaction scores. This approach produces more accurate predictions than standard regression and classification approaches that directly fit the survey scores with call data, and can be easily generalized to other customer satisfaction prediction problems. Source code and data are available at https://github.com/cyberyu/ecml2017.

Twitter provides an open and rich source of data for studying human behaviour at scale and is widely used in social and network sciences. However, a major criticism of Twitter data is that demographic information is largely absent. Enhancing Twitter data with user ages would advance our ability to study social network structures, information flows and the spread of contagions. Approaches toward age detection of Twitter users typically focus on specific properties of tweets, e.g., linguistic features, which are language dependent. In this paper, we devise a language-independent methodology for determining the age of Twitter users from data that is native to the Twitter ecosystem. The key idea is to use a Bayesian framework to generalise ground-truth age information from a few Twitter users to the entire network based on what/whom they follow. Our approach scales to inferring the age of 700 million Twitter accounts with high accuracy.

The World Bank provides billions of dollars in development finance to countries across the world every year. As many projects are related to the environment, we want to understand the World Bank projects impact to forest cover. However, the global extent of these projects results in substantial heterogeneity in impacts due to geographic, cultural, and other factors. Recent research by Athey and Imbens has illustrated the potential for hybrid machine learning and causal inferential techniques which may be able to capture such heterogeneity. We apply their approach using a geolocated dataset of World Bank projects, and augment this data with satellite-retrieved characteristics of their geographic context (including temperature, precipitation, slope, distance to urban areas, and many others). We use this information in conjunction with causal tree (CT) and causal forest (CF) approaches to contrast ‘control’ and ‘treatment’ geographic locations to estimate the impact of World Bank projects on vegetative cover.

This paper studies supervised learning algorithms for the problem of uncooperative direction finding of a radio emitter using the received signal strength indicator (RSSI) from a rotating and uncharacterized antenna. Radio Direction Finding (RDF) is the task of finding the direction of a radio frequency emitter from which the received signal was transmitted, using a single receiver. We study the accuracy of radio direction finding for the 2.4 GHz WiFi band, and restrict ourselves to applying supervised learning algorithms for RSSI information analysis. We designed and built a hardware prototype for data acquisition using off-the-shelf hardware. During the course of our experiments, we collected more than three million RSSI values. We show that we can reliably predict the bearing of the transmitter with an error bounded by 11$$^\circ $$, in both indoor and outdoor environments. We do not explicitly model the multi-path, that inevitably arises in such situations and hence one of the major challenges that we faced in this work is that of automatically compensating for the multi-path and hence the associated noise in the acquired data.

With the rapid growth in smartphone usage, more organizations begin to focus on providing better services for mobile users. User identification can help these organizations to identify their customers and then cater services that have been customized for them. Currently, the use of cookies is the most common form to identify users. However, cookies are not easily transportable (e.g., when a user uses a different login account, cookies do not follow the user). This limitation motivates the need to use behavior biometric for user identification. In this paper, we propose DeepService, a new technique that can identify mobile users based on user’s keystroke information captured by a special keyboard or web browser. Our evaluation results indicate that DeepService is highly accurate in identifying mobile users (over 93% accuracy). The technique is also efficient and only takes less than 1 ms to perform identification.

Transaction frauds impose serious threats onto e-commerce. We present CLUE, a novel deep-learning-based transaction fraud detection system we design and deploy at JD.com, one of the largest e-commerce platforms in China with over 220 million active users. CLUE captures detailed information on users’ click actions using neural-network based embedding, and models sequences of such clicks using the recurrent neural network. Furthermore, CLUE provides application-specific design optimizations including imbalanced learning, real-time detection, and incremental model update. Using real production data for over eight months, we show that CLUE achieves over 3x improvement over the existing fraud detection approaches.

Suspect investigation as a critical function of policing determines the truth about how a crime occurred, as far as it can be found. Understanding of the environmental elements in the causes of a crime incidence inevitably improves the suspect investigation process. Crime pattern theory concludes that offenders, rather than venture into unknown territories, frequently commit opportunistic and serial violent crimes by taking advantage of opportunities they encounter in places they are most familiar with as part of their activity space. In this paper, we present a suspect investigation method, called SINAS, which learns the activity space of offenders using an extended version of the random walk method based on crime pattern theory, and then recommends the top-K potential suspects for a committed crime. Our experiments on a large real-world crime dataset show that SINAS outperforms the baseline suspect investigation methods we used for the experimental evaluation.

In this research, we focus on automatic supervised stance classification of tweets. Given test datasets of tweets from five various topics, we try to classify the stance of the tweet authors as either in FAVOR of the target, AGAINST it, or NONE. We apply eight variants of seven supervised machine learning methods and three filtering methods using the WEKA platform. The macro-average results obtained by our algorithm are significantly better than the state-of-art results reported by the best macro-average results achieved in the SemEval 2016 Task 6-A for all the five released datasets. In contrast to the competitors of the SemEval 2016 Task 6-A, who did not use any char skip ngrams but rather used thousands of ngrams and hundreds of word embedding features, our algorithm uses a few tens of features mainly character-based features where most of them are skip char ngram features.

The growing availability of data from cities (e.g., traffic flow, human mobility and geographical data) open new opportunities for predicting and thus optimizing human activities. For example, the automatic analysis of land use enables the possibility of better administrating a city in terms of resources and provided services. However, such analysis requires specific information, which is often not available for privacy concerns. In this paper, we propose a novel machine learning representation based on the available public information to classify the most predominant land use of an urban area, which is a very common task in urban computing. In particular, in addition to standard feature vectors, we encode geo-social data from Location-Based Social Networks (LBSNs) into a conceptual tree structure that we call Geo-Tree. Then, we use such representation in kernel machines, which can thus perform accurate classification exploiting hierarchical substructure of concepts as features. Our extensive comparative study on the areas of New York and its boroughs shows that Tree Kernels applied to Geo-Trees are very effective improving the state of the art up to 18% in Macro-F1.

Worldwide, conservation agencies employ rangers to protect conservation areas from poachers. However, agencies lack the manpower to have rangers effectively patrol these vast areas frequently. While past work has modeled poachers’ behavior so as to aid rangers in planning future patrols, those models’ predictions were not validated by extensive field tests. In this paper, we present a hybrid spatio-temporal model that predicts poaching threat levels and results from a five-month field test of our model in Uganda’s Queen Elizabeth Protected Area (QEPA). To our knowledge, this is the first time that a predictive model has been evaluated through such an extensive field test in this domain. We present two major contributions. First, our hybrid model consists of two components: (i) an ensemble model which can work with the limited data common to this domain and (ii) a spatio-temporal model to boost the ensemble’s predictions when sufficient data are available. When evaluated on real-world historical data from QEPA, our hybrid model achieves significantly better performance than previous approaches with either temporally-aware dynamic Bayesian networks or an ensemble of spatially-aware models. Second, in collaboration with the Wildlife Conservation Society and Uganda Wildlife Authority, we present results from a five-month controlled experiment where rangers patrolled over 450 sq km across QEPA. We demonstrate that our model successfully predicted (1) where snaring activity would occur and (2) where it would not occur; in areas where we predicted a high rate of snaring activity, rangers found more snares and snared animals than in areas of lower predicted activity. These findings demonstrate that (1) our model’s predictions are selective, (2) our model’s superior laboratory performance extends to the real world, and (3) these predictive models can aid rangers in focusing their efforts to prevent wildlife poaching and save animals.

Signature extraction is a key part of forensic log analysis. It involves recognizing patterns in log lines such that log lines that originated from the same line of code are grouped together. A log signature consists of immutable parts and mutable parts. The immutable parts define the signature, and the mutable parts are typically variable parameter values. In practice, the number of log lines and signatures can be quite large, and the task of detecting and aligning immutable parts of the logs to extract the signatures becomes a significant challenge. We propose a novel method based on a neural language model that outperforms the current state-of-the-art on signature extraction. We use an RNN auto-encoder to create an embedding of the log lines. Log lines embedded in such a way can be clustered to extract the signatures in an unsupervised manner.

The future planning, management and prediction of water demand and usage should be preceded by long-term variation analysis for related parameters in order to enhance the process of developing new scenarios whether for surface-water or ground-water resources. This paper aims to provide an appropriate methodology for long-term prediction for the water flow and water level parameters of the Shannon river in Ireland over a 30-year period from 1983–2013 through a framework that is composed of three phases: city wide scale analytics, data fusion, and domain knowledge data analytics phase which is the main focus of the paper that employs a machine learning model based on deep convolutional neural networks (DeepCNNs). We test our proposed deep learning model on three different water stations across the Shannon river and show it out-performs four well-known time-series forecasting models. We finally show how the proposed model simulate the predicted water flow and water level from 2013–2080. Our proposed solution can be very useful for the water authorities for better planning the future allocation of water resources among competing users such as agriculture, demotic and power stations. In addition, it can be used for capturing abnormalities by setting and comparing thresholds to the predicted water flow and water level.

The rapid growth of Web usage for advertising job positions provides a great opportunity for real-time labour market monitoring. This is the aim of Labour Market Intelligence (LMI), a field that is becoming increasingly relevant to EU Labour Market policies design and evaluation. The analysis of Web job vacancies, indeed, represents a competitive advantage to labour market stakeholders with respect to classical survey-based analyses, as it allows for reducing the time-to-market of the analysis by moving towards a fact-based decision making model. In this paper, we present our approach for automatically classifying million Web job vacancies on a standard taxonomy of occupations. We show how this problem has been expressed in terms of text classification via machine learning. Then, we provide details about the classification pipelines we evaluated and implemented, along with the outcomes of the validation activities. Finally, we discuss how machine learning contributed to the LMI needs of the European Organisation that supported the project.

Interaction networks consist of a static graph with a time-stamped list of edges over which interaction took place. Examples of interaction networks are social networks whose users interact with each other through messages or location-based social networks where people interact by checking in to locations. Previous work on finding influential nodes in such networks mainly concentrate on the static structure imposed by the interactions or are based on fixed models for which parameters are learned using the interactions. In two recent works, however, we proposed an alternative activity data driven approach based on the identification of influence propagation patterns. In the first work, we identify so-called information-channels to model potential pathways for information spread, while the second work exploits how users in a location-based social network check in to locations in order to identify influential locations. To make our algorithms scalable, approximate versions based on sketching techniques from the data streams domain have been developed. Experiments show that in this way it is possible to efficiently find good seed sets for influence propagation in social networks.

Data Cleaning represents a crucial and error prone activity in KDD that might have unpredictable effects on data analytics, affecting the believability of the whole KDD process. In this paper we describe how a bridge between AI Planning and Data Quality communities has been made, by expressing both the data quality and cleaning tasks in terms of AI planning. We also report a real-life application of our approach.

Sequential data can be found in many settings, e.g., as sequences of visited websites or as location sequences of travellers. To improve the understanding of the underlying mechanisms that generate such sequences, the HypTrails approach provides for a novel data analysis method. Based on first-order Markov chain models and Bayesian hypothesis testing, it allows for comparing a set of hypotheses, i.e., beliefs about transitions between states, with respect to their plausibility considering observed data. HypTrails has been successfully employed to study phenomena in the online and the offline world. In this talk, we want to give an introduction to HypTrails and showcase selected real-world applications on urban mobility and reading behavior on Wikipedia.

Finding a parking space is a key problem in urban scenarios, often due to the lack of actual parking availability information for drivers. Modern vehicles, able to identify free parking spaces using standard on-board sensors, have been proven to be effective probes to measure parking availability. Nevertheless, spatio-temporal datasets resulting from probe vehicles pose significant challenges to the machine learning and data mining communities, due to volume, noise, and heterogeneous spatio-temporal coverage. In this paper we summarize some of the approaches we proposed to extract new knowledge from this data, with the final goal to reduce the parking search time. First, we present a spatio-temporal analysis of the suitability of taxi movements for parking crowd-sensing. Second, we describe machine learning approaches to automatically generate maps of parking spots and to predict parking availability. Finally, we discuss some open issues for the ML/KDD community.

In this paper, we summarize our work on using the predictive clustering framework for image analysis. More specifically, we have used predictive clustering trees to generate image representations, that can then be used to perform image retrieval and/or image annotation. We have evaluated the proposed method for performing image retrieval on general purpose images [6], and annotation of general purpose images [5], medical images [3] and diatom images [4].

Music generation has recently become popular as an application of machine learning. To generate polyphonic music, one must consider both simultaneity (the vertical consistency) and sequentiality (the horizontal consistency). Bayesian networks are suitable to model both simultaneity and sequentiality simultaneously. Here, we present music generation models based on Bayesian networks applied to chord voicing, four-part harmonization, and real-time chord prediction.

We describe ProTraits, a machine learning pipeline that systematically annotates microbes with phenotypes using a large amount of textual data from scientific literature and other online resources, as well as genome sequencing data. Moreover, by relying on a multi-view non-negative matrix factorization approach, ProTraits pipeline is also able to discover novel phenotypic concepts from unstructured text. We present the main components of the developed pipeline and outline challenges for the application to other fields.

Process-based modeling is an approach to constructing explanatory models of dynamical systems from knowledge and data. The knowledge encodes information about potential processes that explain the relationships between the observed system entities. The resulting process-based models provide both an explanatory overview of the system components and closed-form equations that allow for simulating the system behavior. In this paper, we present three recent improvements of the process-based approach: (i) improving predictive performance of process-based models using ensembles, (ii) extending the scope of process-based models towards handling uncertainty and (iii) addressing the task of automated process-based design.

Machine-learnt models based on additive ensembles of binary regression trees are currently deemed the best solution to address complex classification, regression, and ranking tasks. Evaluating these models is a computationally demanding task as it needs to traverse thousands of trees with hundreds of nodes each. The cost of traversing such large forests of trees significantly impacts their application to big and stream input data, when the time budget available for each prediction is limited to guarantee a given processing throughput. Document ranking in Web search is a typical example of this challenging scenario, where the exploitation of tree-based models to score query-document pairs, and finally rank lists of documents for each incoming query, is the state-of-art method for ranking (a.k.a. Learning-to-Rank). This paper presents QuickScorer, a novel algorithm for the traversal of huge decision trees ensembles that, thanks to a cache- and CPU-aware design, provides a $${\sim } 9 \! \times $$ speedup over best competitors.

We review our recent progress in the development of graph kernels. We discuss the hash graph kernel framework, which makes the computation of kernels for graphs with vertices and edges annotated with real-valued information feasible for large data sets. Moreover, we summarize our general investigation of the benefits of explicit graph feature maps in comparison to using the kernel trick. Our experimental studies on real-world data sets suggest that explicit feature maps often provide sufficient classification accuracy while being computed more efficiently. Finally, we describe how to construct valid kernels from optimal assignments to obtain new expressive graph kernels. These make use of the kernel trick to establish one-to-one correspondences. We conclude by a discussion of our results and their implication for the future development of graph kernels.

Often the manual review of large data sets, either for purposes of labeling unlabeled instances or for classifying meaningful results from uninteresting (but statistically significant) ones is extremely resource intensive, especially in terms of subject matter expert (SME) time. Use of active learning has been shown to diminish this review time significantly. However, since active learning is an iterative process of learning a classifier based on a small number of SME-provided labels at each iteration, the lack of an enabling tool can hinder the process of adoption of these technologies in real-life, in spite of their labor-saving potential. In this demo we present ASK-the-Expert, an interactive tool that allows SMEs to review instances from a data set and provide labels within a single framework. ASK-the-Expert is powered by an active learning algorithm for training a classifier in the backend. We demonstrate this system in the context of an aviation safety application, but the tool can be adopted to work as a simple review and labeling tool as well, without the use of active learning.

Academic search engines (e.g., Google scholar or Microsoft academic) provide a medium for retrieving various information on scholarly documents. However, most of these popular scholarly search engines overlook the area of data set retrieval, which should provide information on relevant data sets used for academic research. Due to the increasing volume of publications, it has become a challenging task to locate suitable data sets on a particular research area for benchmarking or evaluations. We propose Delve, a web-based system for data set retrieval and document analysis. This system is different from other scholarly search engines as it provides a medium for both data set retrieval and real time visual exploration and analysis of data sets and documents.

Feature selection is an essential step to identify relevant and non-redundant features for target class prediction. In this context, the number of feature combinations grows exponentially with the dimension of the feature space. This hinders the user’s understanding of the feature-target relevance and feature-feature redundancy. We propose an interactive Framework for Exploring and Understanding Multivariate Correlations (FEXUM), which embeds these correlations using a force-directed graph. In contrast to existing work, our framework allows the user to explore the correlated feature space and guides in understanding multivariate correlations through interactive visualizations.

We present an explainable recommendation system for novels and authors, called Lit@EVE, which is based on Wikipedia concept vectors. In this system, each novel or author is treated as a concept whose definition is extracted as a concept vector through the application of an explainable word embedding technique called EVE. Each dimension of the concept vector is labelled as either a Wikipedia article or a Wikipedia category name, making the vector representation readily interpretable. In order to recommend items, the Lit@EVE system uses these vectors to compute similarity scores between a target novel or author and all other candidate items. Finally, the system generates an ordered list of suggested items by showing the most informative features as human-readable labels, thereby making the recommendation explainable.

The paper presents a smartphone application for monitoring physical activity and mental stress. The application utilizes sensor data from a wristband and/or a smartphone, which can be worn in various pockets or in a bag in any orientation. The presence and location of the devices are used as contexts for the selection of appropriate machine-learning models for activity recognition and the estimation of human energy expenditure. The stress-monitoring method uses two machine-learning models, the first one relying solely on physiological sensor data and the second one incorporating the output of the activity monitoring and other context information. The evaluation showed that we recognize a wide range of atomic activities with the accuracy of 87%, and that we outperform the state-of-the art consumer devices in the estimation of energy expenditure. In stress monitoring we achieved the accuracy of 92% in a real-life setting.

Each machine learning task comes equipped with its own set of performance measures. For example, there is a plethora of classification measures that assess predictive performance, a myriad of clustering indices, and equally many rule interestingness measures. Choosing the right measure requires careful thought, as it can influence model selection and thus the performance of the final machine learning system. However, analyzing and understanding measure properties is a difficult task. Here, we present Tetrahedron, a web-based visualization tool that aids the analysis of complete ranges of performance measures based on a two-by-two contingency matrix. The tool operates in a barycentric coordinate system using a 3D tetrahedron, which can be rotated, zoomed, cut, parameterized, and animated. The application is capable of visualizing predefined measures (86 currently), as well as helping prototype new measures by visualizing user-defined formulas.

TF Boosted Trees (TFBT) is a new open-sourced framework for the distributed training of gradient boosted trees. It is based on TensorFlow, and its distinguishing features include a novel architecture, automatic loss differentiation, layer-by-layer boosting that results in smaller ensembles and faster prediction, principled multi-class handling, and a number of regularization techniques to prevent overfitting.

Visualizing frequently occurring patterns and potentially unusual behaviors in trajectory can provide valuable insights into activities behind the data. In this paper, we introduce TrajViz, a motif (frequently repeated subsequences) based visualization software that detects patterns and anomalies by inducing “grammars” from discretized spatial trajectories. We consider patterns as a set of sub-trajectories with unknown lengths that are spatially similar to each other. We demonstrate that TrajViz has the capacity to help users visualize anomalies and patterns effectively.

This paper presents a system for tracking and analyzing the evolution and transformation of topics in an information network. The system consists of four main modules for pre-processing, adaptive topic modeling, network creation and temporal network analysis. The core module is built upon an adaptive topic modeling algorithm adopting a sliding time window technique that enables the discovery of groundbreaking ideas as those topics that evolve rapidly in the network.

We present WHODID: a turnkey intuitive web-based interface for fault detection, identification and diagnosis in production units. Fault detection and identification is an extremely useful feature and is becoming a necessity in modern production units. Moreover, the large deployment of sensors within the stations of a production line has enabled the close monitoring of products being manufactured. In this context, there is a high demand for computer intelligence able to detect and isolate faults inside production lines, and to additionally provide a diagnosis for maintenance on the identified faulty production device, with the purpose of preventing subsequent faults caused by the diagnosed faulty device behavior. We thus introduce a system which has fault detection, isolation, and identification features, for retrospective and on-the-fly monitoring and maintenance of complex dynamical production processes. It provides real-time answers to the questions: “is there a fault?”, “where did it happen?”, “for what reason?”. The method is based on a posteriori analysis of decision sequences in XGBoost tree models, using recurrent neural networks sequential models of tree paths.The particularity of the presented system is that it is robust to missing or faulty sensor measurements, it does not require any modeling of the underlying, possibly exogenous manufacturing process, and provides fault diagnosis along with confidence level in plain English formulations. The latter can be used as maintenance directions by a human operator in charge of production monitoring and control.