Abstract
The ever-growing nature of user generated data in online systems poses obvious challenges on how we process such data. Typically, this issue is regarded as a scalability problem and has been mainly addressed with distributed algorithms able to train on massive amounts of data in short time windows. However, data is inevitably adding up at high speeds. Eventually one needs to discard or archive some of it. Moreover, the dynamic nature of data in user modeling and recommender systems, such as change of user preferences, and the continuous introduction of new users and items make it increasingly difficult to maintain up-to-date, accurate recommendation models. The objective of this workshop is to bring together researchers and practitioners interested in incremental and adaptive approaches to stream-based user modeling, recommendation and personalization, including algorithms, evaluation issues, incremental content and context mining, privacy and transparency, temporal recommendation or software frameworks for continuous learning.

Abstract
Hashtags have become a crucial social media tool. The categorization of posts in a simple and informal way helps to spread the content through the web. At the same time, it enables users to easily find messages within a specific topic. However, the flexibility provided to use and create a hashtag carries some problems. Equivalent expressions, like synonyms, are handled like entirely different words. On the other hand, the same hashtag may refer to different topics. In this paper, we present TORHID (Topic Relevant Hashtag Identification), a method that employs topic modeling with the purpose of retrieving and identifying hashtags relevant to a specific topic in Twitter streams, starting from a seed hashtag and resorting to a classifier to remove non relevant hashtags. The result is a network of hashtags related to the seed, that we can use to deepen the initial search.

Abstract
Online incremental models for recommendation are nowadays pervasive in both the industry and the academia. However, there is not yet a standard evaluation methodology for the algorithms that maintain such models. Moreover, online evaluation methodologies available in the literature generally fall short on the statistical validation of results, since this validation is not trivially applicable to stream-based algorithms. We propose a k-fold validation framework for the pairwise comparison of recommendation algorithms that learn from user feedback streams, using prequential evaluation. Our proposal enables continuous statistical testing on adaptive-size sliding windows over the outcome of the prequential process, allowing practitioners and researchers to make decisions in real time based on solid statistical evidence. We present a set of experiments to gain insights on the sensitivity and robustness of two statistical tests-McNemar's and Wilcoxon signed rank-in a streaming data environment. Our results show that besides allowing a real-time, fine-grained online assessment, the online versions of the statistical tests are at least as robust as the batch versions, and definitely more robust than a simple prequential single-fold approach.

Abstract
Although widely used, the majority of current music recommender systems still focus on recommendations' accuracy, user preferences and isolated item characteristics, without evaluating other important factors, like the joint item selections and the recommendation moment. However, when it comes to playlist recommendations, additional dimensions, as well as the notion of user experience and perception, should be taken into account to improve recommendations' quality. In this work, HybA, a hybrid recommender system for automatic playlist continuation, that combines Latent Dirichlet Allocation and Case-Based Reasoning, is proposed. This system aims to address "similar concepts" rather than similar users. More than generating a playlist based on user requirements, like automatic playlist generation methods, HybA identifies the semantic characteristics of a started playlist and reuses the most similar past ones, to recommend relevant playlist continuations. In addition, support to beyond accuracy dimensions, like increased coherence or diverse items' discovery, is provided. To overcome the semantic gap between music descriptions and user preferences, identify playlist structures and capture songs' similarity, a graph model is used. Experiments on real datasets have shown that the proposed algorithm is able to outperform other state of the art techniques, in terms of accuracy, while balancing between diversity and coherence.

Abstract
Contextual embeddings represent a new generation of semantic representations learned from Neural Language Modelling (NLM) that addresses the issue of meaning conflation hampering traditional word embeddings. In this work, we show that contextual embeddings can be used to achieve unprecedented gains in Word Sense Disambiguation (WSD) tasks. Our approach focuses on creating sense-level embeddings with full-coverage of WordNet, and without recourse to explicit knowledge of sense distributions or task-specific modelling. As a result, a simple Nearest Neighbors (k-NN) method using our representations is able to consistently surpass the performance of previous systems using powerful neural sequencing models. We also analyse the robustness of our approach when ignoring part-of-speech and lemma features, requiring disambiguation against the full sense inventory, and revealing shortcomings to be improved. Finally, we explore applications of our sense embeddings for concept-level analyses of contextual embeddings and their respective NLMs.

Abstract
As the amount of generated information grows, reading and summarizing texts of large collections turns into a challenging task. Many documents do not come with descriptive terms, thus requiring humans to generate keywords on-the-fly. The need to automate this kind of task demands the development of keyword extraction systems with the ability to automatically identify keywords within the text. One approach is to resort to machine-learning algorithms. These, however, depend on large annotated text corpora, which are not always available. An alternative solution is to consider an unsupervised approach. In this article, we describe YAKE!, a light-weight unsupervised automatic keyword extraction method which rests on statistical text features extracted from single documents to select the most relevant keywords of a text. Our system does not need to be trained on a particular set of documents, nor does it depend on dictionaries, external corpora, text size, language, or domain. To demonstrate the merits and significance of YAKE!, we compare it against ten state-of-the-art unsupervised approaches and one supervised method. Experimental results carried out on top of twenty datasets show that YAKE! significantly outperforms other unsupervised methods on texts of different sizes, languages, and domains.