Recommending personalized content to users is a long-standing challenge to many online services including Facebook, Yahoo, Linkedin and Twitter. Traditional recommendation models such as latent factor models and feature-based models are usually trained for all users and optimize an “average” experience for them, yielding sub-optimal solutions. Although multi-task learning provides an opportunity to learn personalized models per user, learning algorithms are usually tailored to specific models (e.g., generalized linear model, matrix factorization and etc.), creating obstacles for a unified engineering interface, which is important for large Internet companies. In this paper, we present an empirical framework to learn user-specific personal models for content recommendation by utilizing gradient information from a global model. Our proposed method can potentially benefit any model that can be optimized through gradients, offering a lightweight yet generic alternative to conventional multi-task learning algorithms for user personalization. We demonstrate the effectiveness of the proposed framework by incorporating it in three popular machine learning algorithms including logistic regression, gradient boosting decision tree and matrix factorization. Our extensive empirical evaluation shows that the proposed framework can significantly improve the efficiency of personalized recommendation in real-world datasets.

Students’ performance in Massive Open Online Courses (MOOCs) is enhanced by high quality discussion forums or recently emerging educational Community Question Answering (CQA) systems. Nevertheless, only a small number of students answer questions asked by their peers. This results in instructor overload, and many unanswered questions. To increase students’ participation, we present an approach for recommendation of new questions to students who are likely to provide answers. Existing approaches to such question routing proposed for non-educational CQA systems tend to rely on a few experts, which is not suitable because we want students to be engaged as it positively influences their learning outcomes. In tackling this novel educational question routing problem, our method (1) goes beyond previous question-answering data as it incorporate additional non-QA data from the course (to improve prediction accuracy and to involve a larger part of community) and (2) applies constraints on users’ workload (to prevent user overloading). We use an ensemble classifier for predicting students’ willingness to answer a question, as well as the students’ expertise for answering. We conducted an online evaluation of the proposed method using an A/B test in our CQA system deployed at an edX MOOC. The proposed method outperformed a baseline method (non-educational question routing enhanced with workload restriction) in recommendation accuracy, involving more community members, and average number of contributions.

The item cold-start problem is inherent to collaborative filtering (CF) recommenders where items and users are represented by vectors in a latent space. It emerges since CF recommenders rely solely on historical user interactions to characterize their item inventory. As a result, an effective serving of new and trendy items to users may be delayed until enough user feedback is received, thus, reducing both users’ and content suppliers’ satisfaction.

To mitigate this problem, many commercial recommenders apply random exploration and devote a small portion of their traffic to explore new items and gather interactions from random users. Alternatively, content or context information is combined into the CF recommender, resulting in a hybrid system. Another hybrid approach is to learn a mapping between the item attribute space and the CF latent feature space, and use it to characterize the new items providing initial estimates for their latent vectors.

In this paper, we adopt the attribute-to-feature mapping approach to expedite random exploration of new items and present LearnAROMA – an advanced algorithm for learning the mapping, previously proposed in the context of classification. In particular, LearnAROMA learns a Gaussian distribution over the mapping matrix. Numerical evaluation demonstrates that this learning technique achieves more accurate initial estimates than logistic regression methods. We then consider a random exploration setting, in which new items are further explored as user interactions arrive. To leverage the initial latent vector estimates with the incoming interactions, we propose DynamicBPR – an algorithm for updating the new item latent vectors without retraining the CF model. Numerical evaluation reveals that DynamicBPR achieves similar accuracy as a CF model trained on all the ratings, using 70% less exploring users than conventional random exploration.

Group recommendation has attracted significant research efforts for its importance in benefiting a group of users. This paper investigates the Group Recommendation problem from a novel aspect which tries to maximize the satisfaction of each group member while minimizing the unfairness between them.

In this work, we present several semantics of the individual utility and propose two concepts of social welfare and fairness for modeling the overall utilities and the balance of group members. We formulate the problem as a multiple objective optimization problem and show its computational complexity (NP-Hardness Analysis) in different semantics. Given the multiple-objective nature of fairness-aware group recommendation problem, we provide an optimization framework for fairness-aware group recommendation from the perspective of Pareto Efficiency. We conduct extensive experiments on real-world datasets (one of which contains real group structures and purchase histories) and evaluate our algorithm with standard accuracy metrics. The results indicate that considering fairness in group recommendation can enhance the recommendation accuracy.

In recommender systems based on low rank factorization of a partially observed user-item matrix, a common phenomenon that plagues many otherwise effective models is the interleaving of good and spurious recommendations in the top-K results. A single incongruous recommendation can dramatically impact the perceived quality of a recommender system. In this work, we investigate folding, a major contributing factor to spurious recommendations. Folding refers to the unintentional overlap of disparate groups of users and items in the low-rank embedding vector space, induced by improper handling of missing data. We formally define a metric that quantifies the severity of folding in a trained system, to assist in diagnosing its potential to shock users with inappropriate recommendations. The folding metric complements existing information retrieval metrics that focus on the number of good recommendations and their ranks but ignore the impact of undesired recommendations. We motivate our definition of the folding metric on synthetic data and evaluate its effectiveness on both synthetic and real world datasets. We study the relationship between the folding metric and other characteristics of recommender systems and observe that optimizing for goodness metrics can lead to high folding and thus more spurious recommendations.

Recently, many e-commerce websites have encouraged their users to rate shopping items and write review text. This review text information has been very useful for understanding user preferences and item properties and it enhances the capability to make personalized recommendations of these websites. In this paper, we propose to model user preferences and item properties using convolutional neural networks (CNNs) with dual local and global attention, motivated by the superiority of CNNs to extract complex features. By using aggregated review text from a user and aggregated review text for an item, our model can learn the unique features (embedding) of each user and each item. These features are then used to predict ratings. We train these user and item networks jointly and this enables the interaction between users and items in a similar way to matrix factorization. The local attention gives us insight on a user’s preferences or an item’s properties. The global attention helps CNNs focus on semantic meanings of the whole review text. Thus, the combined local and global attentions enable an interpretable and better-learned representation of users and items. We validate the proposed models by applying popular review datasets in Yelp and Amazon and compare the results with matrix factorization (MF), the hidden factors as topics (HFT) model, and the recently proposed convolutional matrix factorization (ConvMF+). The proposed CNNs with dual attention model outperforms HFT and ConvMF+ in terms of mean square errors (MSE). In addition, we compare the user/item embeddings learned from these models for classification and recommendation. These results also confirm the superior quality of user/item embeddings learned from our model.

This work addresses the problem of selecting Tensor Factorization algorithms for the Context-aware Filtering recommendation task using a metalearning approach. The most important challenge of applying metalearning on new problems is the development of useful measures able to characterize the data, i.e. metafeatures. We propose an extensive and exhaustive set of metafeatures to characterize Context-aware Filtering recommendation task. These metafeatures take advantage of the tensor’s hierarchical structure via slice operations. The algorithm selection task is addressed as a Label Ranking problem, which ranks the Tensor Factorization algorithms according to their expected performance, rather than simply selecting the algorithm that is expected to obtain the best performance. A comprehensive experimental work is conducted on both levels, baselevel and metalevel (Tensor Factorization and Label Ranking, respectively). The results show that the proposed metafeatures lead to metamodels that tend to rank Tensor Factorization algorithms accurately and that the selected algorithms present high recommendation performance.

The recommendation challenge can be posed as the problem of predicting either item ratings or item rankings. The latter approach has proven more effective. Pairwise learning-to-rank techniques have been relatively successful. Hence, they are popularly used for learning recommender model parameters such as those in collaborative filtering (CF) models. The model parameters are learned by optimizing close smooth approximations of the non-smooth information retrieval (IR) metrics such as Mean Area Under ROC curve (AUC).

Targeted campaigns are an alternative to item recommendations for increasing conversion. The user ranking task is referred to as audience retrieval. It is used in targeted campaigns to rank push campaign recipients based on their potential to convert. In this work, we consider the task of efficiently learning a ranking model that provides item recommendations and user rankings simultaneously. We adopt pairwise learning for this task. We refer to our novel approach as multi-objective pairwise ranking (MPR).

We describe our approach and evaluate its performance by experiments.

Developing a real-world recommender system, i.e. for use in large-scale online retail, poses a number of different challenges. Interestingly, only a small part of these challenges are of algorithmic nature, such as how to select the most accurate model for a given use case. Instead, most technical problems usually arise from operational constraints, such as: adaptation to novel use cases; cost and complexity of system maintenance; capability of reusing pre-existing signal and integrating heterogeneous data sources.

In this paper, we describe the system we developed in order to address those constraints at Zalando, which is one of the most popular online fashion retailers in Europe. In particular, we explain how moving from a collaborative filtering approach to a learning-to-rank model helped us to effectively tackle the challenges mentioned above, while improving at the same time the quality of our recommendations. A fairly detailed description of our software architecture is provided, along with an overview of the algorithmic approach. On the other hand, we present some of the offline and online experiments that we ran in order to validate our models.

We propose a novel latent factor model for recommending product size fits {Small, Fit, Large} to customers. Latent factors for customers and products in our model correspond to their physical true size, and are learnt from past product purchase and returns data. The outcome for a customer, product pair is predicted based on the difference between customer and product true sizes, and efficient algorithms are proposed for computing customer and product true size values that minimize two loss function variants. In experiments with Amazon shoe datasets, we show that our latent factor models incorporating personas, and leveraging return codes show a 17-21% AUC improvement compared to baselines. In an online A/B test, recommendations produced by our algorithms show an improvement of 33 basis points in percentage of Fit transactions over control.

Recurrent Neural Networks are powerful tools for modeling sequences. They are flexibly extensible and can incorporate various kinds of information including temporal order. These properties make them well suited for generating sequential recommendations. In this paper, we extend Recurrent Neural Networks by considering unique characteristics of the Recommender Systems domain. One of these characteristics is the explicit notion of the user recommendations are specifically generated for. We show how individual users can be represented in addition to sequences of consumed items in a new type of Gated Recurrent Unit to effectively produce personalized next item recommendations. Offline experiments on two real-world datasets indicate that our extensions clearly improve objective performance when compared to state-of-the-art recommender algorithms and to a conventional Recurrent Neural Network.

Modeling the complex interactions between users and items as well as amongst items themselves is at the core of designing successful recommender systems. One classical setting is predicting users’ personalized sequential behavior (or ‘next-item’ recommendation), where the challenges mainly lie in modeling ‘third-order’ interactions between a user, her previously visited item(s), and the next item to consume. Existing methods typically decompose these higher-order interactions into a combination pairwise relationships, by way of which user preferences (user-item interactions) and sequential patterns (item-item interactions) are captured by separate components. In this paper, we propose a unified method, TransRec, to model such third-order relationships for large-scale sequential prediction. Methodologically, we embed items into a ‘transition space’ where users are modeled as translation vectors operating on item sequences. Empirically, this approach outperforms the state-of-the-art on a wide spectrum of real-world datasets.

The technical barriers for conversing with recommender systems using natural language are vanishing. Already, there are commercial systems that facilitate interactions with an AI agent. For instance, it is possible to say “what should I watch” to an Apple TV remote to get recommendations. In this research, we investigate how users initially interact with a new natural language recommender to deepen our understanding of the range of inputs that these technologies can expect. We deploy a natural language interface to a recommender system, we observe users’ first interactions and follow-up queries, and we measure the differences between speaking- and typing-based interfaces. We employ qualitative methods to derive a categorization of users’ first queries (objective, subjective, and navigation) and follow-up queries (refine, reformulate, start over). We employ quantitative methods to determine the differences between speech and text, finding that speech inputs are typically longer and more conversational.

We describe a novel application of recommender systems, helping marathon runners to run a new personal-best race-time, by predicting a challenging, but achievable, target-time and by recommending a race-plan to achieve this time that is tailored to their ability and the course. A comprehensive evaluation of prediction accuracy and race-plan quality is provided using a large-scale dataset with almost 400,000 runners from the last 12 years of the Chicago marathon.

Factorization Machines (FMs) are a model class capable of learning pairwise (and in general higher order) feature interactions from high dimensional, sparse data. In this paper we adopt an elementary view on FMs. Specifically, we view FMs as a sum of simple surfaces – a hyperplane plus several squared hyperplanes – in the original feature space; this elementary view, although equivalent to that of low rank matrix factorization, is geometrically more intuitive and points to some interesting generalizations. Led by our intuition, we challenge our understanding of the inductive bias of FMs by showing a simple dataset where FMs counterintuitively fail to learn the weight of the interaction between two features. We discuss the reasons and mathematically formulate and prove a form of this limitation. Also inspired by our elementary view, we propose modeling intermediate orders of interaction, such as 1.5-way FMs. Beyond the specific proposals, the goal of this paper is to expose our thoughts and ideas to the research community, in an effort to take FMs to the next level.

This paper presents recommendation algorithms for novel applications in chemistry. We focus on the task of recommending to the chemist candidate molecules (reactants) necessary to synthesize a given target molecule (product), which is an important step for the chemist to find a synthesis route to generate the product. We formulate this task as a link-prediction problem over a so-called Network of Organic Chemistry (NOC) that we have constructed from 8 million chemical reactions described in the US patent literature between 1976 and 2013. We leverage state-of-the-art factorization algorithms for recommender systems to solve this task. Our empirical evaluation demonstrates that Factorization Machines, trained with chemistry-specific knowledge, outperforms current methods based on similarity of chemical structures.

Research into recommendation algorithms has made great strides in recent years. However, these algorithms are typically applied in relatively straightforward scenarios: given information about a user’s past preferences, what will they like in the future? Recommendation is often more complex, however: evaluating recommended items never takes place in a vacuum, and it is often a single step in the user’s more complex background task. In this paper, we define a specific type of recommendation scenario called narrative-driven recommendation, where the recommendation process is driven by both a log of the user’s past transactions as well as a narrative description of their current interest(s). Through an analysis of a set of real-world recommendation narratives from the LibraryThing forums, we demonstrate the uniqueness and richness of this scenario and highlight common patterns and properties of such narratives.

Knowledge Graphs have proven to be extremely valuable to recommender systems, as they enable hybrid graph-based recommendation models encompassing both collaborative and content information. Leveraging this wealth of heterogeneous information for top-N item recommendation is a challenging task, as it requires the ability of effectively encoding a diversity of semantic relations and connectivity patterns.

In this work, we propose a novel approach to learning user-item relatedness from knowledge graphs for top-N item recommendations. We start from a knowledge graph modeling user-item and item-item relations and we learn property-specific vector representations of users and items applying neural language models on the network. These representations are used to create property-specific user-item relatedness features, which are in turn fed into learning to rank algorithms to learn a global relatedness model that optimizes top-N item recommendations. We evaluate the proposed approach in terms of ranking quality on the MovieLens 1M dataset, outperforming two state-of-the-art recommender systems, and we assess the importance of property-specific relatedness scores on the overall ranking quality.

In the last years, we have seen much attention given to the semantic gap problem in multimedia recommender systems. Much effort has been devoted to bridge this gap by building tools for the extraction of high-level, semantics-based features from multimedia content, as low-level features are not considered useful because they deal primarily with representing the perceived content rather than the semantics of it.

In this paper, we explore a different point of view, by leveraging the gap between low-level and high-level features. We experiment with a recent approach for movie recommendations that extract low-level mise-en-scene features from multimedia content and combine it with high-level features provided by the wisdom of the crowd.

For our purposes, we first designed an empirical study involving 100 subjects and implemented a movie recommender system with three different versions of the same algorithm, respectively based on (i) conventional movie attributes, (ii) mise-en-scene features, and (iii) a combination of mise-en-scene features and movie attributes. We collected data regarding the quality perceived by the users. In another study, we compared users’ perceived utility of recommendation with offline quality measures.

Results from both studies show that the introduction of mise-en-scene features in conjunction with traditional attributes improves both offline and online quality of recommendations.

We propose an encoder-decoder neural architecture to model user session and intent using browsing and purchasing data from a large e-commerce company.

We begin by identifying the source-target transition pairs between items within each session. Then, the set of source items are passed through an encoder, whose learned representation is used by the decoder to estimate the sequence of target items. Therefore, as this process is performed pair-wise, we hypothesize that the model could capture the transition regularities in a more fine grained way. Additionally, our model incorporates an attention mechanism to explicitly learn the more expressive portions of the sequences in order to improve performance. Besides modeling the user sessions, we also extended the original architecture by means of attaching a second decoder that is jointly trained to predict, for each user, the intent of her next session. With this, we want to explore to what extent the model can capture inter session dependencies.

We performed an empirical study comparing against several baselines on a large real world dataset, showing that our approach is competitive in both item and intent prediction.

Users of ecommerce portals commonly use customer reviews for making purchase decisions. Many products contain tens or hundreds of reviews which makes it impossible for the customer to read all of them in order to get a good idea about the product. A review recommendation system that can recommend a subset of the reviews is thus useful for e-commerce websites. However customer reviews are of varied quality and different reviews cover different aspects or issues about the product. We follow previous work that maintain the statistical distribution of product aspects along with their associated sentiments of the entire review set for that particular product. However we address the challenge which arises due to the fact that similar aspects are mentioned in different reviews using different natural language expressions (e.g. camera, photo and picture refers to the same product aspect of camera). We use vector representations to identify mentions of similar aspects and group them together under single product aspect. Review helpfulness score may act as a proxy for quality of reviews, but this approach suffers from cold start problem as new reviews do not have any helpfulness score. We measure the quality of the review based on its content by doing supervised training using a convolutional neural network. Though our neural network model is trained with reviews from Amazon dataset which have helpfulness score, our model can be used to predict the quality score of new reviews which do not have any helpfulness score. The recommended subset of reviews have high content score and the coverage of the product aspects, issues and sentiments are representative of the entire review set. The system is evaluated on datasets from Amazon and is found to be more useful than the competing methods.

Hybrid recommender systems combine several different sources of information to generate recommendations. These systems have been shown to improve accuracy compared to single-source recommendation strategies. However, hybrid recommendation strategies are inherently more complex than those that use a single source of information, and thus the process of explaining recommendations to consumers becomes more challenging. In this paper we show how to use a hybrid recommender system built on a probabilistic programming language to create hybrid recommendations. We then describe an investigation into user preferences for explanations in this system. We present results of an online user survey that evaluates explanations for hybrid algorithms in a variety of text and visual, graph-based formats.

Deep learning methods have led to substantial progress in various application fields of AI, and in recent years a number of proposals were made to improve recommender systems with artificial neural networks. For the problem of making session-based recommendations, i.e., for recommending the next item in an anonymous session, Hidasi et al. recently investigated the application of recurrent neural networks with Gated Recurrent Units (GRU4REC). Assessing the true value of such novel approaches based only on what is reported in the literature is however difficult when no standard evaluation protocols are applied and when the strength of the baselines used in the performance comparison is not clear. In this work we show based on a comprehensive empirical evaluation that a heuristics-based nearest neighbor (kNN) scheme for sessions outperforms GRU4REC in the large majority of the tested configurations and datasets. Neighborhood sampling and efficient in-memory data structures ensure the scalability of the kNN method. The best results in the end were often achieved when we combine the kNN approach with GRU4REC, which shows that RNNs can leverage sequential signals in the data that cannot be detected by the co-occurrence-based kNN method.

We present a research tool for user preference elicitation that collects both explicit user feedback and unobtrusively acquired facial expressions. The concrete implementation is a web-based user interface where the user is presented with two music excerpts. After listening to both, the user provides a pairwise score (i.e. which of the two items is preferred) for each pair of music excerpts. The novelty of the demo is the integration of the unobtrusive acquisition of facial expressions through the webcam. During the listening of the music excerpts, the system extracts features related to the facial expressions of the user several times per second. The interaction runs as a web application, which allows for a large-scale remote acquisition of emotional data. Up to now, such acquisitions were usually done in controlled environments with few subjects, hence being of little use for the recommender systems community.

Typically, the task of managing the social media presence of a company or a public person is the job of a dedicated social media account manager. While many attempts have been made in recent years to provide more automation to account managers, complete workflow automation has still to be achieved. Pokedem is a social media management application that aims at filling this gap, recommending actions that could be performed by the account manager to increase the popularity of a Twitter account and the engagement of its audience. By casting the problem in the setting of recommendation systems, Pokedem is able to provide account managers with a complete tool for automating their daily activities.

We present Citolytics – a novel link-based recommendation system for Wikipedia articles. In a preliminary study, Citolytics achieved promising results compared to the widely used text-based approach of Apache Lucene’s MoreLikeThis (MLT). In this demo paper, we describe how we plan to integrate Citolytics into the Wikipedia infrastructure by using Elasticsearch and Apache Flink to serve recommendations for Wikipedia articles. Additionally, we propose a large-scale online evaluation design using the Wikipedia Android app. Working with Wikipedia data has several unique advantages. First, the availability of a very large user sample contributes to statistically significant results. Second, the openness of Wikipedia’s architecture allows making our source code and evaluation data public, thus benefiting other researchers. If link-based recommendations show promise in our online evaluation, a deployment of the presented system within Wikipedia would have a far-reaching impact on Wikipedia’s more than 30 million users.

Explainable recommendation is an important task. Many methods have been proposed which generate explanations from the content and reviews written for items. When review text is unavailable, generating explanations is still a hard problem. In this paper, we illustrate how explanations can be generated in such a scenario by leveraging external knowledge in the form of knowledge graphs. Our method jointly ranks items and knowledge graph entities using a Personalized PageRank procedure to produce recommendations together with their explanations.

Given a set of candidate items, Recommendation by Explanation constructs a justification for recommending each item, in the form of what we call an Explanation Chain, and then recommends those candidates that have the best explanations. By unifying recommendation and explanation, this approach enables us to find relevant recommendations with explanations that have a high degree of both fidelity and interpretability. Experimental results on a movie recommendation dataset show that our approach also provides sets of recommendations that have a high degree of serendipity, low popularity-bias and high diversity.

Music recommender systems often operate in sequential mode by suggesting a collection of songs that constitute a listening session. This task is usually called automated music playlist generation and it has been previously studied in the literature with different successful approaches based on, e.g., variations of collaborative filtering or content-based similarity. Some of the proposed playlist models take into consideration the current song and a number of previous songs, i.e., the song context , in order to predict the next song. However, it is not yet clear to what extent knowing this song context improves next-song predictions. To shed light on this question, we conduct a numerical experiment on two datasets of hand-curated music playlists, where we compare playlist models that account for different song context lengths. Our results indicate that knowing the song context seems, at first, uninformative. However, we explain this effect by a strong bias in the data towards very popular songs and observe that, in fact, songs in the long tail are more accurately predicted when the song context is considered.

Several studies indicate that attracting students to research careers requires to engage them from early undergraduate years. Following this, the Engineering School at PUC Chile has developed an undergraduate research program that allows students to enroll in research in exchange for course credits. Moreover, we developed a web portal to inform students about the program, but participation remains lower than expected. In order to promote student engagement, we attempt to build a personalized recommender system of research opportunities to undergraduates. With this goal in mind we investigate two tasks. First, identifying students that are more willing to participate on this kind of program. A second task is generating a personalized list of recommendations of research opportunities for each student. To evaluate our approach, we perform a simulated prediction experiment with data from our school, which has more than 4,000 active undergraduate students nowadays. Results indicate that there is a big potential to create a personalized recommender system for this purpose. Our research can be used as a baseline for colleges seeking strategies to encourage research activities within undergraduate students.

Traditionally, recommender systems exploit user ratings to infer preferences. However, the growing popularity of social platforms has encouraged users to write textual reviews about liked items. These reviews represent a valuable source of non-trivial information that could improve users’ decision processes. In this paper we propose a novel recommendation approach based on the semantic annotation of entities mentioned in user reviews and on the knowledge available in the Web of Data. We compared our recommender system with two baseline algorithms and a state-of-the-art Linked Data based approach. Our system provided more diverse recommendations with respect to the other techniques considered, while obtaining a better accuracy than the Linked Data based method.

Many students change their majors during college. Choosing the right academic majors, minors, and concentrations within a major in higher education is challenging but crucial in assuring students’ overall academic and career success. Personalized prediction of student success in majors, minors, and concentrations will help students better find the right majors, minors, and concentrations for them, so as to timely achieve their academic goals. In this paper, we present a new recommender application for the academic major/minor/concentration selection problem in the educational domain. The proposed major/minor/concentration recommender system is a collaborative filtering-based recommender based on student grade prediction, and provides a variety of personalized predictions and recommendations on majors, minors, and concentrations. It is a useful tool to guide students, advisors, and administrators during the personalized major/minor/concentration exploration (declare/change) process.

We present an initial examination on the impact text complexity has when incorporated into the recommendation process in community question answering sites. We use Read2Vec, a readability assessment tool designed to measure the readability level of short documents, to inform a traditional content-based recommendation strategy. The results highlight the benefits of incorporating readability information in this process.

The vast volumes of open data pose a challenge for users in finding relevant datasets. To address this, we developed a hybrid dataset recommendation model that combines content-based similarity with item-to-item co-occurrence. The features used by the recommender include dataset properties and usage statistics. In this paper, we focus on fi ne-tuning the weights of these features. We experimentally compare two feature weighting approaches: a uniform one with predefined weights and a user-driven one, where the weights are informed by the opinions of system users. We evaluated the two approaches in a study, involving the users of a real-life data portal. The results suggest that user-driven feature weights can improve dataset recommendations, although not at all levels of data relevance, and highlight the importance of incorporating target users in the design of recommender systems.

Music emotion recognition (MER) is a key issue in user context-aware recommendation. Many existing methods require hand-crafted features on audio and lyrics. Here we propose a new end-to-end method for recognizing emotions of tracks from their acoustic signals and lyrics via multimodal deep neural networks. We evaluate our method on about 7,000 K-pop tracks labeled as positive or negative emotion. The proposed method is compared to end-to-end unimodal models using audio signals or lyrics only. The experimental results show that our multimodal model achieves the best accuracy as 80%, and we discuss the reasons of these results.

Typical recommender evaluations treat users as a homogeneous unit. However, user subgroups often differ in their tastes, which can result more broadly in diverse recommender needs. Thus, these groups may have different degrees of satisfaction with the provided recommendations. We explore the offline top- N performance of collaborative filtering algorithms across two domains. We find that several strategies achieve higher accuracy for dominant demographic groups, thus increasing the overall performance for the strategy, without providing increased benefits for other users.

Personalization in recommender systems has typically been applied to the underlying algorithms. In contrast, the presentation of individual recommendations—specifically, the various ways in which it can be adapted to suit the user’s needs in a more effective manner—has received relatively little attention by comparison. We present the results of an exploratory survey about users’ choices regarding hotel recommendations and draw preliminary conclusions about whether these choices can influence the presentation of recommendations.