On many online platforms, users can engage with millions of pieces content, which they discover either organically or through algorithmically-generated recommendations. While the short-term benefits recommender systems are well-known, their long-term impacts less well understood. In this work, we study user experience on Spotify, a popular music streaming service, lens diversity—the coherence set songs listens to. We use high-fidelity embedding based listening behavior Spotify to quantify...

Recommender systems play an important role in providing engaging experience on online music streaming services. However, the musical domain presents distinctive challenges to recommender systems: tracks are short, listened multiple times, typically consumed sessions with other tracks, and relevance is highly context-dependent. In this paper, we argue that modeling users' preferences at beginning of a session practical effective way address these challenges. Using dataset from Spotify,...

Algorithmic recommendations shape music consumption at scale, and understanding the impact of various algorithmic models on how content is consumed a central question for streaming platforms. The ability to shift towards less popular different from user's typical historic tastes not only affords platform ways handling issues such as filter bubbles popularity bias, but also contributes maintaining healthy sustainable patterns necessary overall success.

Increasingly, recommender systems are tasked with improving users' long-term satisfaction. In this context, we study a content exploration task, which formalize as bandit problem delayed rewards. There is an apparent trade-off in choosing the learning signal: waiting for full reward to become available might take several weeks, slowing rate of learning, whereas using short-term proxy rewards reflects actual goal only imperfectly. First, develop predictive model that incorporates all...

Recommender systems are a ubiquitous feature of online platforms. Increasingly, they explicitly tasked with increasing users' long-term satisfaction. In this context, we study content exploration task, which formalize as multi-armed bandit problem delayed rewards. We observe that there is an apparent trade-off in choosing the learning signal: Waiting for full reward to become available might take several weeks, hurting rate at happens, whereas measuring short-term proxy rewards reflects...

We consider the problem of predicting users' preferences on online platforms. build recent findings suggesting that change over time, and helping users expand their horizons is important in ensuring they stay engaged. Most existing models user attempt to capture simultaneous preferences: "Users who like A tend B as well". In this paper, we argue these fail anticipate changing preferences. To overcome issue, seek understand structure underlies evolution end, propose Preference Transition...

More stringent requirements for the protection of environment coupled with new incentives materials recovery, lead modern waste management practice on line a more differentiated approach. Separation, or precisely, non-mixing at source, is one most promising strategies. However, before deciding which categories urban should be collected separately, it useful to have detailed knowledge regarding characteristics waste. A 5-year investigation has produced enough information answer such questions...

Inspired by applications in sports where the skill of players or teams competing against each other varies over time, we propose a probabilistic model pairwise-comparison outcomes that can capture wide range time dynamics. We achieve this replacing static parameters class popular models continuous-time Gaussian processes; covariance function these processes enables expressive develop an efficient inference algorithm computes approximate Bayesian posterior distribution. Despite flexbility our...

Online controlled experiments, in which different variants of a product are compared based on an Overall Evaluation Criterion (OEC), have emerged as gold standard for decision making online services. It is vital that the OEC aligned with overall goal stakeholders effective making. However, this challenge when not immediately observable. For instance, we might want to understand effect deploying feature long-term retention, where outcome (retention) observable at end A/B test.

Digital media platforms give users access to enormous amounts of content that they must explore avoid boredom and satisfy their needs for heterogeneity. Existing strands work across psychology, marketing, computer science, music underscore the importance lifecycle understanding exploratory behavior, but are also often inconsistent with each other. In this study, we examine how online on Spotify over time, whether by discovering entirely novel or refreshing listening habits from one time...

Knowledge of waste composition is crucial importance for management forecasting. Composition usually specified by average content glass, paper, organic matter etc. In this paper a sorting method and its application to variance determination described. A variation coefficient confidence interval are then calculated. From these two parameters an appreciation the dispersion uncertainty associated with mean values can be derived. case studied, coefficients calculated were between 0.10 0.50...

We study the problem of optimizing a recommender system for outcomes that occur over several weeks or months. begin by drawing on reinforcement learning to formulate comprehensive model users' recurring relationships with system. Measurement, attribution, and coordination challenges complicate algorithm design. describe careful modeling -- including new representation user state key conditional independence assumptions which overcomes these leads simple, testable prototypes. apply our...

On Spotify, multiple recommender systems enable personalized user experiences across a wide range of product features. These are owned by different teams and serve goals, but all these need to explore learn about new content as it appears on the platform. In this work, we describe ongoing efforts at Spotify develop an efficient solution problem, centralizing exploration providing signals existing, decentralized recommendation (a.k.a. exploitation systems). We take creator-centric...

In this work, we draw attention to a connection between skill-based models of game outcomes and Gaussian process classification models. The perspective enables a) principled way dealing with uncertainty b) rich models, specified through kernel functions. Using connection, tackle the problem predicting football matches national teams. We develop player that relates any two players lined up on field. This makes it possible share knowledge gained from observing clubs (available in large...

We address the problem of learning a ranking by using adaptively chosen pairwise comparisons. Our goal is to recover accurately but sample comparisons sparingly. If all comparison outcomes are consistent with ranking, optimal solution use an efficient sorting algorithm, such as Quicksort. But how do algorithms behave if some inconsistent ranking? give favorable guarantees for Quicksort popular Bradley-Terry model, under natural assumptions on parameters. Furthermore, we empirically...

As the number of contributors to online peer-production systems grows, it becomes increasingly important predict whether edits that users make will eventually be beneficial project. Existing solutions either rely on a user reputation system or consist highly specialized predictor is tailored specific system. In this work, we explore different point in solution space goes beyond but does not involve any content-based feature edits. We view each edit as game between editor and component posit...

We investigate the role of uncertainty in decision-making problems with natural language as input. For such tasks, using Large Language Models agents has become norm. However, none recent approaches employ any additional phase for estimating agent about world during task. focus on a fundamental framework input, which is one contextual bandits, where context information consists text. As representative no estimation, we consider an LLM bandit greedy policy, picks action corresponding to...