We consider the problem of generating diverse, personalized recommendations such that a small set of recommended items covers a broad range of the user’s interests. We represent items in a similarity graph, and we formulate the relevance/diversity trade-off as finding a small set of unrated items that best covers a subset of items positively rated by the user. In contrast to previous approaches, our method does not rely on an explicit trade-off between a relevance objective and a diversity objective, as the estimations of relevance and diversity are implicit in the coverage criterion. We show on several benchmark datasets that our approach compares favorably to the state-of-the-art diversification methods according to various relevance and diversity measures.

Full text in ACM Digital Library
Paper Session 1 – Beyond Accuracy

Determinantal point processes (DPPs) are an emerging model for encoding probabilities over subsets, such as shopping baskets, selected from a ground set, such as an item catalog. They have recently proved to be appealing models for a number of machine learning tasks, including product recommendation. DPPs are parametrized by a positive semi-definite kernel matrix. Prior work has shown that using a low-rank factorization of this kernel provides scalability improvements that open the door to training on large-scale datasets and computing online recommendations, both of which are infeasible with standard DPP models that use a full-rank kernel. A low-rank DPP model can be trained using an optimization-based method, such as stochastic gradient ascent, to find a point estimate of the kernel parameters, which can be performed efficiently on large-scale datasets. However, this approach requires careful tuning of regularization parameters to prevent overfitting and provide good predictive performance, which can be computationally expensive. In this paper we present a Bayesian method for learning a low-rank factorization of this kernel, which provides automatic control of regularization. We show that our Bayesian low-rank DPP model can be trained efficiently using stochastic gradient Hamiltonian Monte Carlo (SGHMC). Our Bayesian model generally provides better predictive performance on several real-world product recommendation datasets than optimization-based low-rank DPP models trained using stochastic gradient ascent, and better performance than several state-of-the art recommendation methods in many cases.

Full text in ACM Digital Library
Paper Session 11 – Algorithms II

Explanations are important for users to make decisions on whether to take recommendations. However, algorithm generated explanations can be overly simplistic and unconvincing. We believe that humans can overcome these limitations. Inspired by how people explain word-of-mouth recommendations, we designed a process, combining crowdsourcing and computation, that generates personalized natural language explanations. We modeled key topical aspects of movies, asked crowdworkers to write explanations based on quotes from online movie reviews, and personalized the explanations presented to users based on their rating history. We evaluated the explanations by surveying 220 MovieLens users, finding that compared to personalized tag-based explanations, natural language explanations: 1) contain a more appropriate amount of information, 2) earn more trust from users, and 3) make users more satisfied. This paper contributes to the research literature by describing a scalable process for generating high quality and personalized natural language explanations, improving on state-of-the-art content-based explanations, and showing the feasibility and advantages of approaches that combine human wisdom with algorithmic processes.

Full text in ACM Digital Library
Paper Session 5 – Trust and Reliability

Discovering what people are known for is valuable to many important applications such as recommender systems. Unlike an individual’s personal interests, what a user is known for is reflected by the views of others, and is often not easily discerned for a long-tail of the vast majority of users. In this paper, we tackle the problem of discovering what users are known for through a probabilistic model called Bayesian Contextual Poisson Factorization. Moving beyond just modeling user’s content, it naturally models and integrates additional contextual factors, concretely, user’s geo-spatial footprints and social influence, to overcome noisy online activities and social relations. Through GPS-tagged social media datasets, we find that the proposed method can improve known-for prediction performance by 17.5% in precision and 20.9% in recall on average, and that it can capture the implicit relationships between a user’s known-for profile and her content, geo-spatial and social influence.

Full text in ACM Digital Library
Paper Session 9 – Contextual Challenges

Short-length random walks on the bipartite user-item graph have recently been shown to provide accurate and diverse recommendations. Nonetheless, these approaches suffer from severe time and space requirements, which can be alleviated via random walk sampling, at the cost of reduced recommendation quality. In addition, these approaches ignore users’ ratings, which further limits their expressiveness. In this paper, we introduce a computationally efficient graph-based approach for collaborative filtering based on short-path enumeration. Moreover, we propose three scoring functions based on the Bayesian paradigm that effectively exploit distributional aspects of the users’ ratings. We experiment with seven publicly available datasets against state-of-the-art graph-based and matrix factorization approaches. Our empirical results demonstrate the effectiveness of the proposed approach, with significant improvements in most settings. Furthermore, analytical results demonstrate its efficiency compared to other graph-based approaches.

Full text in ACM Digital Library
Paper Session 11 – Algorithms II

Click-through rate (CTR) prediction plays an important role in computational advertising. Models based on degree-2 polynomial mappings and factorization machines (FMs) are widely used for this task. Recently, a variant of FMs, field-aware factorization machines (FFMs), outperforms existing models in some world-wide CTR-prediction competitions. Based on our experiences in winning two of them, in this paper we establish FFMs as an effective method for classifying large sparse data including those from CTR prediction. First, we propose efficient implementations for training FFMs. Then we comprehensively analyze FFMs and compare this approach with competing models. Experiments show that FFMs are very useful for certain classification problems. Finally, we have released a package of FFMs for public use.

Full text in ACM Digital Library
Paper Session 2 – Algorithms I

As users browse a recommender system, they systematically consider or skip over much of the displayed content. It seems obvious that these eye gaze patterns contain a rich signal concerning these users’ preferences. However, because eye tracking data is not available to most recommender systems, these signals are not widely incorporated into personalization models. In this work, we show that it is possible to predict gaze by combining easily-collected user browsing data with eye tracking data from a small number of users in a grid-based recommender interface. Our technique is able to leverage a small amount of eye tracking data to infer gaze patterns for other users. We evaluate our prediction models in MovieLens — an online movie recommender system. Our results show that incorporating eye tracking data from a small number of users significantly boosts accuracy as compared with only using browsing data, even though the eye-tracked users are different from the testing users (e.g. AUC=0.823 vs. 0.693 in predicting whether a user will fixate on an item). We also demonstrate that Hidden Markov Models (HMMs) can be applied in this setting; they are better than linear models in predicting fixation probability and capturing the interface regularity through Bayesian inference (AUC=0.823 vs. 0.757).

Full text in ACM Digital Library
Paper Session 4 – User in the Loop

An accurate and comprehensive user modeling technique is crucial for the quality of recommender systems. Traditionally, we model user preferences using only actions from the target site and may suffer from cold-start problem. As nowadays people normally engage in multiple online sites for various needs, we consider leveraging the cross-site actions to improve the user modeling accuracy. Specifically, in this paper we aim at achieving a more comprehensive and accurate user modeling by modeling user’s actions in multiple aligned heterogeneous sites simultaneously. To do so, we propose a modularized probabilistic graphical model framework JUMA. We further integrate topic model and matrix factorization into JUMA for joint user modeling over text-based and item-based sites. We assemble and publish large-scale dataset for comprehensive analyzing and evaluation. Experimental results show that our framework JUMA out performs traditional within-site user modeling techniques, especially for cold-start scenarios. For cold-start users, we achieve relative improvements of 9.3% and 12.8% comparing to existing within-site approaches for recommendation in item-based and text-based sites respectively. Thus we draw the conclusion that aligning heterogeneous sites and modeling users jointly do help to improve the quality of online recommender systems.

Full text in ACM Digital Library
Paper Session 3 – Cold Start and Hybrid Methods

Existing feature-based recommendation methods incorporate auxiliary features about users and/or items to address data sparsity and cold start issues. They mainly consider features that are organized in a flat structure, where features are independent and in a same level. However, auxiliary features are often organized in rich knowledge structures (e.g. hierarchy) to describe their relationships. In this paper, we propose a novel matrix factorization framework with recursive regularization — ReMF, which jointly models and learns the influence of hierarchically-organized features on user-item interactions, thus to improve recommendation accuracy. It also provides characterization of how different features in the hierarchy co-influence the modeling of user-item interactions. Empirical results on real-world data sets demonstrate that ReMF consistently outperforms state-of-the-art feature-based recommendation methods.

Full text in ACM Digital Library
Paper Session 2 – Algorithms I

Strategic behaviour from sellers on e-commerce websites, such as faking transactions and manipulating the recommendation scores through artificial reviews, have been among the most notorious obstacles that prevent websites from maximizing the efficiency of their recommendations. Previous approaches have focused almost exclusively on machine learning-related techniques to detect and penalize such behaviour. In this paper, we tackle the problem from a different perspective, using the approach of the field of mechanism design. We put forward a game model tailored for the setting at hand and aim to construct truthful mechanisms, i.e. mechanisms that do not provide incentives for dishonest reputation-augmenting actions, that guarantee good recommendations in the worst-case. For the setting with two agents, we propose a truthful mechanism that is optimal in terms of social efficiency. For the general case of m agents, we prove both lower and upper bound results on the effciency of truthful mechanisms and propose truthful mechanisms that yield significantly better results, when compared to an existing mechanism from a leading e-commerce site on real data.

Full text in ACM Digital Library
Paper Session 5 – Trust and Reliability

This work is motivated by the need to provide reliable information recommendation to users in social sensing. Social sensing has become an emerging application paradigm that uses humans as sensors to observe and report events in the physical world. These human sensed observations are often viewed as binary claims (either true or false). A fundamental challenge in social sensing is how to ascertain the credibility of claims and the reliability of sources without knowing either of them a priori. We refer to this challenge as truth discovery. While prior works have made progress on addressing this challenge, an important limitation exists: they did not explore the mood sensitivity aspect of the problem. Therefore, the claims identified as correct by current solutions can be completely biased in regards to the mood of human sources and lead to useless or even misleading recommendations. In this paper, we present a new analytical model that explicitly considers the mood sensitivity feature in the solution of truth discovery problem. The new model solves a multi-dimensional estimation problem to jointly estimate the correctness and mood neutrality of claims as well as the reliability and mood sensitivity of sources. We compare our model with state-of-the-art truth discovery solutions using four real world datasets collected from Twitter during recent disastrous and emergent events: Brussels Bombing, Paris Attack, Oregon Shooting, Baltimore Riots, which occurred in 2015 and 2016. The results show that our model has significant improvements over the compared baselines by finding more correct and mood neutral claims.

Full text in ACM Digital Library
Paper Session 5 – Trust and Reliability

Improving the performance of recommender systems using knowledge graphs is an important task. There have been many hybrid systems proposed in the past that use a mix of content-based and collaborative filtering techniques to boost the performance. More recently, some work has focused on recommendations that use external knowledge graphs (KGs) to supplement content-based recommendation. In this paper, we investigate three methods for making KG based recommendations using a general-purpose probabilistic logic system called ProPPR. The simplest of the models, EntitySim, uses only the links of the graph. We then extend the model to TypeSim that also uses the types of the entities to boost its generalization capabilities. Next, we develop a graph based latent factor model, GraphLF, which combines the strengths of latent factorization with graphs. We compare our approaches to a recently proposed state-of-the-art graph recommendation method on two large datasets, Yelp and MovieLens-100K. The experiments illustrate that our approaches can give large performance improvements. Additionally, we demonstrate that knowledge graphs give maximum advantage when the dataset is sparse, and gradually become redundant as more training data becomes available, and hence are most useful in cold-start settings.

Full text in ACM Digital Library
Paper Session 11 – Algorithms II

Users express their personal preferences through ratings, adoptions, and other consumption behaviors. We seek to learn latent representations for user preferences from such behavioral data. One representation learning model that has been shown to be effective for large preference datasets is Restricted Boltzmann Machine (RBM). While homophily, or the tendency of friends to share their preferences at some level, is an established notion in sociology, thus far it has not yet been clearly demonstrated on RBM-based preference models. The question lies in how to appropriately incorporate social network into the architecture of RBM-based models for learning representations of preferences. In this paper, we propose two potential architectures: one that models social network among users as additional observations, and another that incorporates social network into the sharing of hidden units among related users. We study the efficacies of these proposed architectures on publicly available, real-life preference datasets with social networks, yielding useful insights.

Full text in ACM Digital Library
Paper Session 10 – Social Perspective

We address the challenge of personalized recommendation of high quality content producers in social media. While some candidates are easily identifiable (say, by being “favorited” many times), there is a long-tail of potential candidates for whom we have little evidence. Through careful modeling of contextual factors like the geo-spatial, topical, and social preferences of users, we propose a tensor-based personalized expert recommendation framework that integrates these factors for revealing latent connections between homogeneous entities (e.g., users and users) and between heterogeneous entities (e.g., users and experts). Through extensive experiments over geo-tagged Twitter data, we find that the proposed framework can improve the quality of recommendation by over 30% in both precision and recall compared to the state-of-the-art.

Full text in ACM Digital Library
Paper Session 9 – Contextual Challenges

Understanding users’ interactions with highly subjective content—like artistic images—is challenging due to the complex semantics that guide our preferences. On the one hand one has to overcome `standard’ recommender systems challenges, such as dealing with large, sparse, and long-tailed datasets. On the other, several new challenges present themselves, such as the need to model content in terms of its visual appearance, or even social dynamics, such as a preference toward a particular artist that is independent of the art they create. In this paper we build large-scale recommender systems to model the dynamics of a vibrant digital art community, Behance, consisting of tens of millions of interactions (clicks and `appreciates’) of users toward digital art. Methodologically, our main contributions are to model (a) rich content, especially in terms of its visual appearance; (b) temporal dynamics, in terms of how users prefer `visually consistent’ content within and across sessions; and (c) social dynamics, in terms of how users exhibit preferences both towards certain art styles, as well as the artists themselves.

Full text in ACM Digital Library
Paper Session 10 – Social Perspective

Classical recommender systems provide users with ranked lists of recommendations, where each one consists of a single item. However, these ranked lists are not suitable for applications such as trip planning, which deal with heterogeneous items. In this paper, we focus on the problem of recommending a set of packages to the user, where each package is constituted with a set of different Points of Interest that may constitute a tour. Given a collection of POIs, our goal is to recommend the most interesting packages for the user, where each package satisfies the budget constraints. We formally define the problem and we present a novel composite recommendation system, inspired from composite retrieval. Experimental evaluation of our proposed system, using a real-world dataset demonstrates its quality and its ability to improve both diversity and relevance of recommendations.

Full text in ACM Digital Library
Paper Session 6 – Applications

Computing useful recommendations for cold-start users is a major challenge in the design of recommender systems, and additional data is often required to compensate the scarcity of user feedback. In this paper we address such problem in a target domain by exploiting user preferences from a related auxiliary domain. Following a rigorous methodology for cold-start, we evaluate a number of recommendation methods on a dataset with positive-only feedback in the movie and music domains, both in single and cross-domain scenarios. Comparing the methods in terms of item ranking accuracy, diversity and catalog coverage, we show that cross-domain preference data is useful to provide more accurate suggestions when user feedback in the target domain is scarce or not available at all, and may lead to more diverse recommendations depending on the target domain. Moreover, evaluating the impact of the user profile size and diversity in the source domain, we show that, in general, the quality of target recommendations increases with the size of the profile, but may deteriorate with too diverse profiles.

Full text in ACM Digital Library
Paper Session 3 – Cold Start and Hybrid Methods

The cold start problem arises in various recommendation applications. In this paper, we propose a tensor factorization-based algorithm that exploits content features extracted from music audio to deal with the cold start problem for the emerging application next-song recommendation. Specifically, the new algorithm learns sequential behavior to predict the next song that a user would be interested in based on the last song the user just listened to. A unique characteristic of the algorithm is that it learns and updates the mapping between the audio feature space and the item latent space each time during the iterations of the factorization process. This way, the content features can be better exploited in forming the latent features for both users and items, leading to more effective solutions for cold-start recommendation. Evaluation on a large-scale music recommendation dataset shows that the recommendation result of the proposed algorithm exhibits not only higher accuracy but also better novelty and diversity, suggesting its applicability in helping a user explore new items in next-item recommendation. Our implementation is available at https://github.com/fearofchou/ALMM.

Full text in ACM Digital Library
Paper Session 3 – Cold Start and Hybrid Methods

We introduce an asynchronous distributed stochastic gradient algorithm for matrix factorization based collaborative filtering. The main idea of this approach is to distribute the user-rating matrix across different machines, each having access only to a part of the information, and to asynchronously propagate the updates of the stochastic gradient optimization across the network. Each time a machine receives a parameter vector, it averages its current parameter vector with the received one, and continues its iterations from this new point. Additionally, we introduce a similarity based regularization that constrains the user and item factors to be close to the average factors of their similar users and items found on subparts of the distributed user-rating matrix. We analyze the impact of the regularization terms on MovieLens (100K, 1M, 10M) and NetFlix datasets and show that it leads to a more efficient matrix factorization in terms of Root Mean Square Error (RMSE) and Mean Absolute Error (MAE), and that the asynchronous distributed approach significantly improves in convergence time as compared to an equivalent synchronous distributed approach.

Full text in ACM Digital Library
Paper Session 2 – Algorithms I

Most evaluations of novel algorithmic contributions assess their accuracy in predicting what was withheld in an offline evaluation scenario. However, several doubts have been raised that standard offline evaluation practices are not appropriate to select the best algorithm for field deployment. The goal of this work is therefore to compare the offline and the online evaluation methodology with the same study participants, i.e. a within users experimental design. This paper presents empirical evidence that the ranking of algorithms based on offline accuracy measurements clearly contradicts the results from the online study with the same set of users. Thus the external validity of the most commonly applied evaluation methodology is not guaranteed.

Full text in ACM Digital Library
Paper Session 1 – Beyond Accuracy

Prior work relevant to incorporating personality into recommender systems falls into two categories: social science studies and algorithmic ones. Social science studies of preference have found only small relationships between personality and category preferences, whereas, algorithmic approaches found a little improvement when incorporating personality into recommendations. As a result, despite good reasons to believe personality assessments should be useful in recommenders, we are left with no substantial demonstrated impact. In this work, we start with user data from a live recommender system, but study category-by-category variations in preference (both rating levels and distribution) across different personality types. By doing this, we hope to isolate specific areas where personality is most likely to provide value in recommender systems, while also modeling an analytic process that can be used in other domains. After controlling for the family-wise error rate, we find that High Agreeableness users rate at least 0.5 stars higher on a 5-star scale compared to low Agreeableness users. We also find differences in consumption in four different personality types between people who manifested high and low levels of that personality.

Full text in ACM Digital Library
Paper Session 4 – User in the Loop

The intent-aware diversification framework was introduced initially in information retrieval and adopted to the context of recommender systems in the work of Vargas et al. The framework considers a set of aspects associated with items to be recommended. For instance, aspects may correspond to genres in movie recommendations. The framework depends on input aspect model consisting of item selection or relevance probabilities, given an aspect, and user intents, in the form of probabilities that the user is interested in each aspect. In this paper, we examine a number of input aspect models and evaluate the impact that different models have on the framework. In particular, we propose a constrained PLSA model that allows for interpretable output, in terms of known aspects, while achieving greater performance that the explicit co-occurrence counting method used in previous work. We evaluate the proposed models using a well-known MovieLens dataset for which item genres are available.

Full text in ACM Digital Library
Paper Session 1 – Beyond Accuracy

Preparing recommendations for unknown users or such that correctly respond to the short-term needs of a particular user is one of the fundamental problems for e-commerce. Most of the common Recommender Systems assume that user identification must be explicit. In this paper a Session-Aware Recommender System approach is presented where no straightforward user information is required. The recommendation process is based only on user activity within a single session, defined as a sequence of events. This information is incorporated in the recommendation process by explicit context modeling with factorization methods and a novel approach with Recurrent Neural Network (RNN). Compared to the session modeling approach, RNN directly models the dependency of user observed sequential behavior throughout its recurrent structure. The evaluation discusses the results based on sessions from real-life system with ephemeral items (identified only by the set of their attributes) for the task of top-n best recommendations.

Full text in ACM Digital Library
Paper Session 9 – Contextual Challenges

Most research on group recommender systems relies on the assumption that individuals have conflicting preferences; in order to generate group recommendations the system should identify a fair way of aggregating these preferences. Both empirical studies and theoretical frameworks have tried to identify the most effective preference aggregation techniques without coming to definite conclusions. In this paper, we propose to approach group recommendation from the group dynamics perspective and analyze the group decision making process for a particular task (in the travel domain). We observe several individual and group properties and correlate them to choice satisfaction. Supported by these initial results we therefore advocate for the development of new group recommendation techniques that consider group dynamics and support the full group decision making process.

Full text in ACM Digital Library
Paper Session 4 – User in the Loop

Many recommendation techniques rely on the knowledge of preferences data in the form of ratings for items. In this paper, we focus on pairwise preferences as an alternative way for acquiring user preferences and building recommendations. In our scenario, users provide pairwise preference scores for a set of item pairs, indicating how much one item in each pair is preferred to the other. We propose a matrix factorization (MF) and a nearest neighbor (NN) prediction techniques for pairwise preference scores. Our MF solution maps users and items pairs to a joint latent features vector space, while the proposed NN algorithm leverages specific user-to-user similarity functions well suited for comparing users preferences of that type. We compare our approaches to state of the art solutions and show that our solutions produce more accurate pairwise preferences and ranking predictions.

Full text in ACM Digital Library
Paper Session 4 – User in the Loop

Repeat Purchases have become increasingly important in measuring customer’s satisfaction and loyalty to e-commerce websites in regard to online shopping. In this paper, we first propose a model for estimating repeat purchase frequency in a given time period from a given product category using Poisson/Gamma model. Second, we estimate the purchase probabilities of different product types in a product category for each customer using Dirichlet model. Experimental results on data collected by a real-world e-commerce website show that it can predict a user’s average repeat purchase frequency along with their product types with decent accuracy. We also argue that the output of our models can be used as prior information to enhance the performance of time-sensitive recommendation.

Full text in ACM Digital Library
Paper Session 11 – Algorithms II

In this position paper, we take the experimental approach of putting algorithms aside, and reflect on what recommenders would be for people if they were not tied to technology. By looking at some of the shortcomings that current recommenders have fallen into and discussing their limitations from a human point of view, we ask the question: if freed from all limitations, what should, and what could, RecSys be? We then turn to the idea that life itself is the best recommender system, and that people themselves are the query. By looking at how life brings people in contact with options that suit their needs or match their preferences, we hope to shed further light on what current RecSys could be doing better. Finally, we look at the forms that RecSys could take in the future. By formulating our vision beyond the reach of usual considerations and current limitations, including business models, algorithms, data sets, and evaluation methodologies, we attempt to arrive at fresh conclusions that may inspire the next steps taken by the community of researchers working on RecSys.

Full text in ACM Digital Library
Paper Session 7 – Past, Present, Future

How can you discover something new, that matches your interest? Recommender Systems have been studied since the 90ies. Their benefit comes from guiding a user through the density of the information jungle to useful knowledge clearings. Early research on recommender systems focuses on algorithms and their evaluation to improve recommendation accuracy using F-measures and other methodologies from signal-detection theory. Present research includes other aspects such as human factors that affect the user experience and interactive visualization techniques to support transparency of results and user control. In this paper, we analyze all publications on recommender systems from the scopus database, and particularly also papers with such an HCI focus. Based on an analysis of these papers, future topics for recommender systems research are identified, which include more advanced support for user control, adaptive interfaces, affective computing and applications in high risk domains.

Full text in ACM Digital Library
Paper Session 4 – User in the Loop

When the Netflix Prize launched in 2006, it put a spotlight on the importance and use of recommender systems in real-world applications. The competition provided many lessons, and many more have been learned since the Grand Prize was awarded in 2009. The use of recommender systems in industry has continued to grow driven by the availability of many kinds of user data and the continued interest for the area within the research community. In this paper, we will describe what we see as the past, present, and future of recommender systems from an industry perspective.

Full text in ACM Digital Library
Paper Session 7 – Past, Present, Future

Every day, we are confronted with an abundance of decisions that require us to choose from a seemingly endless number of choice options. Recommender systems are supposed to help us deal with this formidable task, but some scholars claim that these systems instead put us inside a “Filter Bubble” that severely limits our perspectives. This paper presents a new direction for recommender systems research with the main goal of supporting users in developing, exploring, and understanding their unique personal preferences.

Full text in ACM Digital Library
Paper Session 1 – Beyond Accuracy

A critical change has occurred in the status of context in recommender systems. In the past, context has been considered `additional evidence’. This past picture is at odds with many present application domains, where user and item information is scarce. Such domains face continuous cold start conditions and must exploit session rather than user information. In this paper, we describe the `Contextual Turn’: the move towards context-driven recommendation algorithms for which context is critical, rather than additional. We cover application domains, algorithms that promise to address the challenges of context-driven recommendation, and the steps that the community has taken to tackle context-driven problems. Our goal is to point out the commonalities of context-driven problems, and urge the community to address the overarching challenges that context-driven recommendation poses.

Full text in ACM Digital Library
Paper Session 9 – Contextual Challenges

Collaboration platforms have been available for many years, but they struggle to overcome what is known as the collaboration paradox: people want all the benefits provided by such tools, but they won’t invest their times without some sort of immediate gain. To tackle this problem, we developed an enterprise collaboration platform named Smart Canvas. Smart Canvas is a service that allows people to bring, create and curate content relevant to their organizations. The product’s main goal is to leverage companies knowledge and tear down silos by connecting people, teams, and content. One of the main components of the product is a scalable cloud-based Recommender System, built using techniques like Topic Modeling, Content-Based Recommenders and Graph Analysis.

In this demo, Smart Canvas will be presented in web and mobile interfaces, with data from a real customer. We are also going to describe details of the models, algorithms and architecture (which includes tools like Apache Spark and Titan Graph Database deployed on Google Cloud Platform).

Full text in ACM Digital Library

Outbrain is the Web’s leading content discovery service, recommending billions of stories daily to a global audience across many of the world’s most prestigious and respected publishers. Outbrain’s recommendation technology combines contextual cues with personalization, where the personalization aspects are a combination of content-based and collaborative filtering techniques.

This paper, and the accompanying demo of a production system, offer a behind-the-scenes view of the content-based aspects of Outbrain’s personalization technology. We detail the types of features we extract from content, as well as the attributes we keep in each user’s content-affinity profile. We then describe and demonstrate how we update each user’s profile, in real time, as the user consumes content while browsing the Web.

Full text in ACM Digital Library

With the ubiquity of fast internet connections and the growing availability of Video-On-Demand (VOD) services powerful recommender systems are needed. Traditionally, movie recommender systems apply user-based collaborative filtering providing high quality recommendations if users maintain user profiles describing preferences and movie ratings. The shortcomings of Collaborative Filtering are that comprehensive user profiles are required and users tend to get recommendations very similar to the user profile (filter bubble). In addition, CF-based recommenders neither consider current trends nor the context. In order to overcome these weaknesses, we develop a system identifying interesting events in the stream of current news and deploying this information for computing recommendations. Our system gathers topics of interest from Twitter and RSS-Feeds, extracts relevant Named Entities, and uses semantic relations for recommending movies closely related to these topics. We explain the used algorithms and show that our system provides highly relevant recommendations.

Full text in ACM Digital Library

The present work proposes a novel approach for a future shopping system. Customers’ personal data are protected by a blockchain-based storage network. Based on the bitcoin protocol, the system transacts encrypted data in a tamper-proof way and is able to run secure multiparty computations while no one but the data owner has access to the input data. Thus, a potential customer is able to allow a company to apply functions like a recommendation algorithm without revealing personal data. In combination with a low-energy transmitter (beacon), a completely new shopping experience arises. The beacon automatically triggers a recommendation process based on encrypted personal data. The resulting outcome is a recommendation system, a self-checkout system, and a payment system all in one, thereby full anonymity is guaranteed and the customer never lose control on her data.

We consider how dynamic A/B experiments — used to discover and deploy policies for personalizing items — can be interpreted from the perspective of a recommendation problem, where no prior data is available. We present an illustrative data set we collected, to evaluate algorithms for recommending emails (that vary along dimensions like subject lines) that will maximize response rate, from different subgroups of online learners. This problem is formalized as a contextual bandit, and do an offline regret comparison of how standard bandit algorithms would perform in optimizing response rate. We report a system that provides an API for real–time data exchange and recommendation policy updates from algorithms from external machine learning researchers, and compare our best-performing offline algorithm — Thompson Sampling — against a randomized policy.

Context-aware systems enable the sensing and analysis of user context in order to provide personalized services to users. We observed that it is possible to automatically learn contextual factors and behavioral patterns when users interact with the system. We later utilize the learned patterns to infer contextual user interests within a recommender system. We present a novel context-aware model for detecting users’ preferred items’ categories using an unsupervised deep learning technique applied to mobile sensor data. We train an auto-encoder for each item genre, using contextual data that was obtained when users interacted with the system. Given new contextual sensor data from a user, the discovered patterns from each auto-encoder are used to predict the category of items that should be recommended to the user in the given context. In order to collect rich contextual data, we conducted an extensive field study over a period of four weeks with a group of ninety users. The analysis reveals significant insights regarding the inference of different granularity levels of categories that are available within the data.

This paper proposes a novel framework for automatic dish discovery via word embeddings on restaurant reviews. We collect a dataset of user reviews from Yelp and parse the reviews to extract dish words. Then, we utilize the processed reviews as training texts to learn the embedding vectors of words via the skip-gram model. In the paper, a nearest- neighbor like score function is proposed to rank the dishes based on their learned representations. We brief some analyses on the preliminary experiments and present a web-based visualization at http://clip.csie.org/yelp/.

In this paper we present a time-based genre prediction strategy that can inform the book recommendation process. To explicitly consider time in predicting genres of interest, we rely on a popular time series forecasting model as well as reading patterns of each individual reader or group of readers (in case of libraries or publishing companies). Based on a conducted initial assessment using the Amazon dataset, we demonstrate our strategy outperforms its baseline counterpart.

In real-world scenarios, recommenders face non-functional requirements of technical nature and must handle dynamic data in the form of sequential streams. Evaluation of recommender systems must take these issues into account in order to be maximally informative. In this paper, we present a framework called Idomaar which enables the efficient multi-dimensional benchmarking of recommender algorithms. Idomaar goes beyond current academic research practices by creating a realistic evaluation environment and computing both effectiveness and technical metrics for stream-based as well as set based evaluation. The potentials of the framework are illustrated in a scenario that focuses on the “research to prototyping to productization” cycle at a company. We show that Idomaar simplifies the testing with different configurations and supports the flexible integration of different data.

Many Collaborative Filtering (CF) algorithms are item-based in the sense that they analyze item-item relations in order to produce item similarities. Recently, several works in the field of Natural Language Processing (NLP) suggested to learn a latent representation of words using neural embedding algorithms. Among them, the Skip-gram with Negative Sampling (SGNS), also known as Word2vec, was shown to provide state-of-the-art results on various linguistics tasks. In this paper, we show that item-based CF can be cast in the same framework of neural word embedding. Inspired by SGNS, we describe a method we name Item2vec for item-based CF that produces embedding for items in a latent space. The method is capable of inferring item-item relations even when user information is not available. We present experimental results that demonstrate the effectiveness of the Item2vec method and show it is competitive with SVD.

Neural embedding techniques are being applied in a growing number of machine learning applications. In this work, we demonstrate a neural embedding technique to model users’ session activity. Specifically, we consider a dataset collected from Microsoft’s App Store consisting of user sessions that include sequential click actions and item purchases. Our goal is to learn a latent manifold that captures users’ session activity and can be utilized for contextual recommendations in an online app store.

Music playlists usually contain some particular musical styles or atmospheres in which users would like to be involved. Music streaming services, such as Spotify, Apple Music, and KKBOX, even allow users to edit and listen to playlists online. While it has been some well-known methods that can nicely model the preference between users and songs, little has been done in the literature to recommend music playlists, each of which can be considered as a set of many individual songs, to users. In the light of this, this paper proposes a preference embedding based on a user-song-playlist graph to learn the preference representations of these three entities. After the embedding process, we then use the learned representations to perform the task of playlist recommendation. Experiments conducted on a real-world dataset show that the proposed embedding method outperforms the baseline of popularity; in addition, we also make a comparison with DeepWalk and LINE for the recommendation task, and the results show that the proposed method can stand comparison with the two state-of-the-art graph embedding techniques.

In order to generate effective results, it is essential for a recommender system to model the information about the user interests in a profile. Even though word embeddings (i.e., vector representations of textual descriptions) have proven to be effective in many contexts, a content-based recommendation approach that employs them is still less effective than collaborative strategies (e.g., SVD). In order to overcome this issue, we introduce a novel criterion to evaluate the word-embedding representation of the items a user rated. The proposed approach defines a vector space in which the similarity between an unevaluated item and those in a user profile is measured in terms of linear independence. Experiments show its effectiveness to perform a better ranking of the items, w.r.t. collaborative filtering, both when compared to a latent-factor-based approach (SVD) and to a classic neighborhood user-based system.

Foursquare is a local search and discovery app where as part of the experience users leave tips, short reviews and suggestions to help other users find great places. This poster summarizes the strategy we use to select the best tips for a given venue. Our new ranking model to select the tips that provide our users with the most informative and high quality content. The introduction of the new model yielded significant and positive results in our metrics from various applications within the Foursquare app ecosystem.

The quality of recommendations is known to be affected by diversity and novelty in addition to accuracy. Recent work has focused on methods that increase diversity of recommendation lists. However, these methods assume the user preference for diversity is constant across all users. In this paper, we show that users’ propensity towards diversity varies greatly and argue that the diversity of recommendation lists should be consistent with the level of user interest in diverse recommendations. We introduce a user segmentation approach in order to personalize recommendation according to user preference for diversity. We show that recommendations generated using these segments match the diversity preferences of users in each segment. We also discuss the impact of this segmentation on the novelty of recommendations.