Natural Language Understanding (NLU) models on voice-controlled speakers face several challenges. In particular, music streaming services have large catalogs, often containing millions of songs, artists, and albums and several thousands of custom playlists and stations. In many cases there is ambiguity and little structural difference between carrier phrases and entity names. In this work, we describe how we leveraged multi-armed bandits in combination with implicit customer feedback to improve accuracy and personalization of responses to voice request in the music domain. Our models are tested in a large-scale industrial system containing several other components. In particular, we focused on using this technology to correct errors made by upstream NLU models and personalize responses based on customer preferences and music provider functionality. The models resulted in significant improvement of playback rate for Amazon Music and are deployed in systems serving several countries and languages. We further used the implicit feedback of the customers to generate weakly labeled training data for the NLU models. This improved the experience for customers using other music providers on all Alexa devices.

Success of manufacturing companies largely depends on reliability of their products. Scheduled maintenance is widely used to ensure that equipment is operating correctly so as to avoid unexpected breakdowns. Such maintenance is often carried out separately for every component, based on its usage or simply on some fixed schedule. However, scheduled maintenance is labor-intensive and ineffective in identifying problems that develop between technician's visits. Unforeseen failures still frequently occur. In contrast, predictive maintenance techniques help determine the condition of in-service equipment in order to predict when and what repairs should be performed. The main goal of predictive maintenance is to enable pro-active scheduling of corrective work, and thus prevent unexpected equipment failures.

Mining frequent patterns is plagued by the problem of pattern explosion, making pattern reduction techniques a key challenge in pattern mining. In this article we propose a novel theoretical framework for pattern reduction by measuring the robustness of a property of an itemset such as closedness or nonderivability. The robustness of a property is the probability that this property holds on random subsets of the original data. We study four properties, namely an itemset being closed, free, non-derivable, or totally shattered, and demonstrate how to compute the robustness analytically without actually sampling the data. Our concept of robustness has many advantages: Unlike statistical approaches for reducing patterns, we do not assume a null hypothesis or any noise model and, in contrast to noise-tolerant or approximate patterns, the robust patterns for a given property are always a subset of the patterns with this property. If the underlying property is monotonic then the measure is also monotonic, allowing us to efficiently mine robust itemsets. We further derive a parameter-free technique for ranking itemsets that can be used for top- k approaches. Our experiments demonstrate that we can successfully use the robustness measure to reduce the number of patterns and that ranking yields interesting itemsets.

Improving the performance of classifiers using pattern mining techniques has been an active topic of data mining research. In this work we introduce the recent temporal pattern mining framework for finding predictive patterns for monitoring and event detection problems in complex multivariate time series data. This framework first converts time series into time-interval sequences of temporal abstractions. It then constructs more complex temporal patterns backwards in time using temporal operators. We apply our framework to health care data of 13,558 diabetic patients and show its benefits by efficiently finding useful patterns for detecting and diagnosing adverse medical conditions that are associated with diabetes.

Compression based pattern mining has been successfully applied to many data mining tasks. We propose an approach based on the minimum description length principle to extract sequential patterns that compress a database of sequences well. We show that mining compressing patterns is NP-Hard and belongs to the class of inapproximable problems. We propose two heuristic algorithms to mining compressing patterns. The first uses a two-phase approach similar to Krimp for itemset data. To overcome performance with the required candidate generation we propose GoKrimp, an effective greedy algorithm that directly mines compressing patterns. We conduct an empirical study on six real-life datasets to compare the proposed algorithms by run time, compressibility, and classification accuracy using the patterns found as features for SVM classifiers.

Mining frequent patterns is plagued by the problem of pattern explosion making pattern reduction techniques a key challenge in pattern mining. In this paper we propose a novel theoretical framework for pattern reduction. We do this by measuring the robustness of a property of an item set such as closed ness or non-derivability. The robustness of a property is the probability that this property holds on random subsets of the original data. We study four properties: closed, free, non-derivable and totally shattered item sets, demonstrating how we can compute the robustness analytically without actually sampling the data. Our concept of robustness has many advantages: Unlike statistical approaches for reducing patterns, we do not assume a null hypothesis or any noise model and the patterns reported are simply a subset of all patterns with this property as opposed to approximate patterns for which the property does not really hold. If the underlying property is monotonic, then the measure is also monotonic, allowing us to efficiently mine robust item sets. We further derive a parameter-free technique for ranking item sets that can be used for top-k approaches. Our experiments demonstrate that we can successfully use the robustness measure to reduce the number of patterns and that ranking yields interesting itemsets.

We present a new approach to mining patterns from symbolic interval data that extends previous approaches by allowing semi-intervals and partially ordered patterns. The mining algorithm combines and adapts efficient algorithms from sequential pattern and itemset mining for discovery of the new semi-interval patterns. The semi-interval patterns and semi-interval partial order patterns are more flexible than patterns over full intervals, and are empirically demonstrated to be more useful as features in classification settings. We performed an extensive empirical evaluation on seven real life interval databases totalling over 146k intervals from more than 400 classes demonstrating the flexibility and usefulness of the patterns.

We present a unifying view of temporal concepts and data models in order to categorize existing approaches for unsupervised pattern mining from symbolic temporal data. We distinguish time point-based methods and interval-based methods as well as univariate and multivariate methods. For each of the main categories we present an algorithm and experimental results in detail. For time points, sequential pattern mining algorithms can be used to express equality and order of time points with gaps in multivariate data. Recently, efficient algorithms have been proposed to mine the more general concept of partial order from time points. For time interval data with precise start and end points the relations of Allen can be used to formulate patterns. The recently proposed Time Series Knowledge Representation is more robust on noisy data and offers an alternative semantic that avoids ambiguity and is more expressive. The overall goal of the tutorial is to show the audience which temporal data models can be used to represent observations of dynamic processes and how patterns can be mined from them.

We present a new visualization of the distance and cluster structure of high dimensional data. It is particularly well suited for analysis tasks of users unfamiliar with complex data analysis techniques as it builds on the well known concept of pie charts. The non-linear projection capabilities of Emergent Self-Organizing Maps (ESOM) are used to generate a topology-preserving ordering of the data points on a circle. The distance structure within the high dimensional space is visualized on the circle analogously to the U-Matrix method for two-dimensional SOM. The resulting display resembles pie charts but has an organic structure that naturally emerges from the data. Pie segments correspond to groups of similar data points. Boundaries between segments represent low density regions with larger distances among neighboring points in the high dimensional space. The representation of distances in the form of a periodic sequence of values makes time series segmentation applicable to automated clustering of the data that is in sync with the visualization. We discuss the usefulness of the method on a variety of data sets to demonstrate the applicability in applications such as document analysis or customer segmentation.

Cover song detection is becoming a very hot research topic when plentiful personal music recordings or performance are released on the Internet. A nice cover song recognizer helps us group and detect cover songs to improve the searching experience. The traditional detection is to match two musical audio sequences by exhaustive pairwise comparisons. Different from the existing work, our aim is to generate a group of concatenated feature sets based on regression modeling and arrange them by indexing-based approximate techniques to avoid complicated audio sequence comparisons. We mainly focus on using exact locality sensitive mapping (ELSM) to join the concatenated feature sets and soft hash values. Similarity-invariance among audio sequence comparison is applied to define an optimal combination of several audio features. Soft hash values are pre-calculated to help locate searching range more accurately. Furthermore, we implement our algorithms in analyzing the real audio cover songs and grouping and detecting a batch of relevant cover songs embedded in large audio datasets.

The BioJournalMonitor is a decision support system for the analysis of trends and topics in the biomedical literature. Its main goal is to identify potential diagnostic and therapeutic biomarkers for specific diseases. Several data sources are continuously integrated to provide the user with up-to-date information on current research in this field. State-of-the-art text mining technologies are deployed to provide added value on top of the original content, including named entity detection, relation extraction, classification, clustering, ranking, summarization, and visualization. We present two novel technologies that are related to the analysis of temporal dynamics of text archives and associated ontologies. Currently, the MeSH ontology is used to annotate the scientific articles entering the PubMed database with medical terms. Both the maintenance of the ontology as well as the annotation of new articles is performed largely manually. We describe how probabilistic topic models can be used to annotate recent articles with the most likely MeSH terms. This provides our users with a competitive advantage because, when searching for MeSH terms, articles are found long before they are manually annotated. We further present a study on how to predict the inclusion of new terms in the MeSH ontology. The results suggest that early prediction of emerging trends is possible. The trend ranking functions are deployed in our system to enable interactive searches for the hottest new trends relating to a disease.

We develop a content-based audio COver Song IdeNtification (COSIN) system to detect/group cover songs.The COSIN takes music audio content as input and performs similarity searching to locate variants of the input (i.e., cover versions). Identified cover songs are returned in the rank order according to their similarity to the input.The COSIN also incorporates a set of tools to evaluate retrieval performance so researchers can explore different retrieval schemes and parameters (e.g. recall, precision).The COSIN utilizes a suite of techniques to detect cover songs including: Pitch + Dynamic Programming (DP), Chroma + DP, and Semantic Feature Summarization (SFS) + Hash-Based Approximate Matching (HBAM). Demonstration system shows that COSIN is a very potential music content retrieval tool. Running some music retrieval schemes on COSIN platform, recent experiments with SFS + LSH Variants demonstrate a nicely balanced efficiency (search speed) v. performance (search accuracy) tradeoff.

Temporal patterns composed of symbolic intervals are commonly formulated with Allen's interval relations originating in temporal reasoning. This representation has severe disadvantages for knowledge discovery. The Time Series Knowledge Representation (TSKR) is a new hierarchical language for interval patterns expressing the temporal concepts of coincidence and partial order. We present effective and efficient mining algorithms for such patterns based on itemset techniques. A novel form of search space pruning effectively reduces the size of the mining result to ease interpretation and speed up the algorithms. On a real data set a concise set of TSKR patterns can explain the underlying temporal phenomena, whereas the patterns found with Allen's relations are far more numerous yet only explain fragments of the data.

Data mining in large collections of polyphonic music has recently received increasing interest by companies along with the advent of commercial online distribution of music. Important applications include the categorization of songs into genres and the recommendation of songs according to musical similarity and the customer's musical preferences. Modeling genre or timbre of polyphonic music is at the core of these tasks and has been recognized as a difficult problem. Many audio features have been proposed, but they do not provide easily understandable descriptions of music. They do not explain why a genre was chosen or in which way one song is similar to another. We present an approach that combines large scale feature generation with meta learning techniques to obtain meaningful features for musical similarity. We perform exhaustive feature generation based on temporal statistics and train regression models to summarize a subset of these features into a single descriptor of a particular notion of music. Using several such models we produce a concise semantic description of each song. Genre classification models based on these semantic features are shown to be better understandable and almost as accurate as traditional methods.

Timbre distance and similarity are expressions of the phenomenon that some music appears similar while other songs sound very different to us. The notion of genre is often used to categorize music, but songs from a single genre do not necessarily sound similar and vice versa. In this work, we analyze and compare a large amount of different audio features and psychoacoustic variants thereof for the purpose of modeling timbre distance. The sound of polyphonic music is commonly described by extracting audio features on short time windows during which the sound is assumed to be stationary. The resulting down sampled time series are aggregated to form a high-level feature vector describing the music. We generated high-level features by systematically applying static and temporal statistics for aggregation. The temporal structure of features in particular has previously been largely neglected. A novel supervised feature selection method is applied to the huge set of possible features. The distances of the selected feature correspond to timbre differences in music. The features show few redundancies and have high potential for explaining possible clusters. They outperform seven other previously proposed feature sets on several datasets with respect to the separation of the known groups of timbrally different music.

Knowledge Discovery in time series usually requires symbolic time series. Many discretization methods that convert numeric time series to symbolic time series ignore the temporal order of values. This often leads to symbols that do not correspond to states of the process generating the time series and cannot be interpreted meaningfully. We propose a new method for meaningful unsupervised discretization of numeric time series called Persist. The algorithm is based on the Kullback-Leibler divergence between the marginal and the self-transition probability distributions of the discretization symbols. Its performance is evaluated on both artificial and real life data in comparison to the most common discretization methods. Persist achieves significantly higher accuracy than existing static methods and is robust against noise. It also outperforms Hidden Markov Models for all but very simple cases.

The performance of feed-forward neural networks trained with the backpropagation algorithm on a dedicated Beowulf cluster is analyzed. The concept of training set parallelism is applied. A new model for run time and speedup prediction is developed. With the model the speedup and efficiency of one iteration of the neural networks can be estimated as a function of block size and cluster size. The model is applied to three example problems representing different applications and network architectures. The estimation of the model has a higher accuracy than traditional methods for run time estimation and can be efficiently calculated. Experiments show that speedup of one iteration does not necessarily translate to a shorter training time toward a given error level. To overcome this problem a heuristic extension to training set parallelism called weight averaging is developed. The results show that training in parallel should only be done on clusters with high performance network connections or a multiprocessor machine. A rule of thumb is given for how much network performance of the cluster is needed to achieve speedup of the training time for a neural network.