Vladan Radosavljevic

Marco De Nadai, Edoardo D'Amico, Max Lefarov et al.

LLMs are increasingly applied to recommendation, retrieval, and reasoning, yet deploying a single end-to-end model that can jointly support these behaviors over large, heterogeneous catalogs remains challenging. Such systems must generate unambiguous references to real items, handle multiple entity types, and operate under strict latency and reliability constraints requirements that are difficult to satisfy with text-only generation. While tool-augmented recommender systems address parts of this problem, they introduce orchestration complexity and limit end-to-end optimization. We view this setting as an instance of a broader research problem: how to adapt LLMs to reason jointly over multiple-domain entities, users, and language in a fully self-contained manner. To this end, we introduce NEO, a framework that adapts a pre-trained decoder-only LLM into a tool-free, catalog-grounded generator. NEO represents items as SIDs and trains a single model to interleave natural language and typed item identifiers within a shared sequence. Text prompts control the task, target entity type, and output format (IDs, text, or mixed), while constrained decoding guarantees catalog-valid item generation without restricting free-form text. We refer to this instruction-conditioned controllability as language-steerability. We treat SIDs as a distinct modality and study design choices for integrating discrete entity representations into LLMs via staged alignment and instruction tuning. We evaluate NEO at scale on a real-world catalog of over 10M items across multiple media types and discovery tasks, including recommendation, search, and user understanding. In offline experiments, NEO consistently outperforms strong task-specific baselines and exhibits cross-task transfer, demonstrating a practical path toward consolidating large-scale discovery capabilities into a single language-steerable generative model.

Edoardo D'Amico, Marco De Nadai, Praveen Chandar et al.

Podcast listening is often grounded in a set of favorite shows, while listener intent can evolve over time. This combination of stable preferences and changing intent motivates recommendation approaches that support both familiarity and exploration. Traditional recommender systems typically emphasize long-term interaction patterns, and are less explicitly designed to incorporate rich contextual signals or flexible, intent-aware discovery objectives. In this setting, models that can jointly reason over semantics, context, and user state offer a promising direction. Large Language Models (LLMs) provide strong semantic reasoning and contextual conditioning for discovery-oriented recommendation, but deploying them in production introduces challenges in catalog grounding, user-level personalization, and latency-critical serving. We address these challenges with GLIDE, a production-scale generative recommender for podcast discovery at Spotify. GLIDE formulates recommendation as an instruction-following task over a discretized catalog using Semantic IDs, enabling grounded generation over a large inventory. The model conditions on recent listening history and lightweight user context, while injecting long-term user embeddings as soft prompts to capture stable preferences under strict inference constraints. We evaluate GLIDE using offline retrieval metrics, human judgments, and LLM-based evaluation, and validate its impact through large-scale online A/B testing. Across experiments involving millions of users, GLIDE increases non-habitual podcast streaming on Spotify home surface by up to 5.4% and new-show discovery by up to 14.3%, while meeting production cost and latency constraints.

Mihajlo Grbovic, Nemanja Djuric, Vladan Radosavljevic et al.

Sponsored search represents a major source of revenue for web search engines. This popular advertising model brings a unique possibility for advertisers to target users' immediate intent communicated through a search query, usually by displaying their ads alongside organic search results for queries deemed relevant to their products or services. However, due to a large number of unique queries it is challenging for advertisers to identify all such relevant queries. For this reason search engines often provide a service of advanced matching, which automatically finds additional relevant queries for advertisers to bid on. We present a novel advanced matching approach based on the idea of semantic embeddings of queries and ads. The embeddings were learned using a large data set of user search sessions, consisting of search queries, clicked ads and search links, while utilizing contextual information such as dwell time and skipped ads. To address the large-scale nature of our problem, both in terms of data and vocabulary size, we propose a novel distributed algorithm for training of the embeddings. Finally, we present an approach for overcoming a cold-start problem associated with new ads and queries. We report results of editorial evaluation and online tests on actual search traffic. The results show that our approach significantly outperforms baselines in terms of relevance, coverage, and incremental revenue. Lastly, we open-source learned query embeddings to be used by researchers in computational advertising and related fields.

Marco De Nadai, Francesco Fabbri, Paul Gigioli et al.

In the ever-evolving digital audio landscape, Spotify, well-known for its music and talk content, has recently introduced audiobooks to its vast user base. While promising, this move presents significant challenges for personalized recommendations. Unlike music and podcasts, audiobooks, initially available for a fee, cannot be easily skimmed before purchase, posing higher stakes for the relevance of recommendations. Furthermore, introducing a new content type into an existing platform confronts extreme data sparsity, as most users are unfamiliar with this new content type. Lastly, recommending content to millions of users requires the model to react fast and be scalable. To address these challenges, we leverage podcast and music user preferences and introduce 2T-HGNN, a scalable recommendation system comprising Heterogeneous Graph Neural Networks (HGNNs) and a Two Tower (2T) model. This novel approach uncovers nuanced item relationships while ensuring low latency and complexity. We decouple users from the HGNN graph and propose an innovative multi-link neighbor sampler. These choices, together with the 2T component, significantly reduce the complexity of the HGNN model. Empirical evaluations involving millions of users show significant improvement in the quality of personalized recommendations, resulting in a +46% increase in new audiobooks start rate and a +23% boost in streaming rates. Intriguingly, our model's impact extends beyond audiobooks, benefiting established products like podcasts.

Sandro Hauri, Nemanja Djuric, Vladan Radosavljevic et al.

National Basketball Association (NBA) players are highly motivated and skilled experts that solve complex decision making problems at every time point during a game. As a step towards understanding how players make their decisions, we focus on their movement trajectories during games. We propose a method that captures the multi-modal behavior of players, where they might consider multiple trajectories and select the most advantageous one. The method is built on an LSTM-based architecture predicting multiple trajectories and their probabilities, trained by a multi-modal loss function that updates the best trajectories. Experiments on large, fine-grained NBA tracking data show that the proposed method outperforms the state-of-the-art. In addition, the results indicate that the approach generates more realistic trajectories and that it can learn individual playing styles of specific players.

Fang-Chieh Chou, Tsung-Han Lin, Henggang Cui et al.

Following detection and tracking of traffic actors, prediction of their future motion is the next critical component of a self-driving vehicle (SDV) technology, allowing the SDV to operate safely and efficiently in its environment. This is particularly important when it comes to vulnerable road users (VRUs), such as pedestrians and bicyclists. These actors need to be handled with special care due to an increased risk of injury, as well as the fact that their behavior is less predictable than that of motorized actors. To address this issue, in the current study we present a deep learning-based method for predicting VRU movement, where we rasterize high-definition maps and actor's surroundings into a bird's-eye view image used as an input to deep convolutional networks. In addition, we propose a fast architecture suitable for real-time inference, and perform an ablation study of various rasterization approaches to find the optimal choice for accurate prediction. The results strongly indicate benefits of using the proposed approach for motion prediction of VRUs, both in terms of accuracy and latency.

Henggang Cui, Vladan Radosavljevic, Fang-Chieh Chou et al.

Autonomous driving presents one of the largest problems that the robotics and artificial intelligence communities are facing at the moment, both in terms of difficulty and potential societal impact. Self-driving vehicles (SDVs) are expected to prevent road accidents and save millions of lives while improving the livelihood and life quality of many more. However, despite large interest and a number of industry players working in the autonomous domain, there still remains more to be done in order to develop a system capable of operating at a level comparable to best human drivers. One reason for this is high uncertainty of traffic behavior and large number of situations that an SDV may encounter on the roads, making it very difficult to create a fully generalizable system. To ensure safe and efficient operations, an autonomous vehicle is required to account for this uncertainty and to anticipate a multitude of possible behaviors of traffic actors in its surrounding. We address this critical problem and present a method to predict multiple possible trajectories of actors while also estimating their probabilities. The method encodes each actor's surrounding context into a raster image, used as input by deep convolutional networks to automatically derive relevant features for the task. Following extensive offline evaluation and comparison to state-of-the-art baselines, the method was successfully tested on SDVs in closed-course tests.

Nemanja Djuric, Vladan Radosavljevic, Henggang Cui et al.

We address one of the crucial aspects necessary for safe and efficient operations of autonomous vehicles, namely predicting future state of traffic actors in the autonomous vehicle's surroundings. We introduce a deep learning-based approach that takes into account a current world state and produces raster images of each actor's vicinity. The rasters are then used as inputs to deep convolutional models to infer future movement of actors while also accounting for and capturing inherent uncertainty of the prediction task. Extensive experiments on real-world data strongly suggest benefits of the proposed approach. Moreover, following completion of the offline tests the system was successfully tested onboard self-driving vehicles.

Abraham Bagherjeiran, Nemanja Djuric, Mihajlo Grbovic et al.

Proceedings of the 2017 AdKDD and TargetAd Workshop held in conjunction with the 23rd ACM SIGKDD Conference on Knowledge Discovery and Data Mining Halifax, Nova Scotia, Canada.

Nemanja Djuric, Mihajlo Grbovic, Vladan Radosavljevic et al.

We consider the problem of personalization of online services from the viewpoint of ad targeting, where we seek to find the best ad categories to be shown to each user, resulting in improved user experience and increased advertisers' revenue. We propose to address this problem as a task of ranking the ad categories depending on a user's preference, and introduce a novel label ranking approach capable of efficiently learning non-linear, highly accurate models in large-scale settings. Experiments on a real-world advertising data set with more than 3.2 million users show that the proposed algorithm outperforms the existing solutions in terms of both rank loss and top-K retrieval performance, strongly suggesting the benefit of using the proposed model on large-scale ranking problems.

Nemanja Djuric, Hao Wu, Vladan Radosavljevic et al.

We consider the problem of learning distributed representations for documents in data streams. The documents are represented as low-dimensional vectors and are jointly learned with distributed vector representations of word tokens using a hierarchical framework with two embedded neural language models. In particular, we exploit the context of documents in streams and use one of the language models to model the document sequences, and the other to model word sequences within them. The models learn continuous vector representations for both word tokens and documents such that semantically similar documents and words are close in a common vector space. We discuss extensions to our model, which can be applied to personalized recommendation and social relationship mining by adding further user layers to the hierarchy, thus learning user-specific vectors to represent individual preferences. We validated the learned representations on a public movie rating data set from MovieLens, as well as on a large-scale Yahoo News data comprising three months of user activity logs collected on Yahoo servers. The results indicate that the proposed model can learn useful representations of both documents and word tokens, outperforming the current state-of-the-art by a large margin.

Erik Ordentlich, Lee Yang, Andy Feng et al.

Word2vec is a popular family of algorithms for unsupervised training of dense vector representations of words on large text corpuses. The resulting vectors have been shown to capture semantic relationships among their corresponding words, and have shown promise in reducing a number of natural language processing (NLP) tasks to mathematical operations on these vectors. While heretofore applications of word2vec have centered around vocabularies with a few million words, wherein the vocabulary is the set of words for which vectors are simultaneously trained, novel applications are emerging in areas outside of NLP with vocabularies comprising several 100 million words. Existing word2vec training systems are impractical for training such large vocabularies as they either require that the vectors of all vocabulary words be stored in the memory of a single server or suffer unacceptable training latency due to massive network data transfer. In this paper, we present a novel distributed, parallel training system that enables unprecedented practical training of vectors for vocabularies with several 100 million words on a shared cluster of commodity servers, using far less network traffic than the existing solutions. We evaluate the proposed system on a benchmark dataset, showing that the quality of vectors does not degrade relative to non-distributed training. Finally, for several quarters, the system has been deployed for the purpose of matching queries to ads in Gemini, the sponsored search advertising platform at Yahoo, resulting in significant improvement of business metrics.

Mihajlo Grbovic, Vladan Radosavljevic, Nemanja Djuric et al.

As one of the leading platforms for creative content, Tumblr offers advertisers a unique way of creating brand identity. Advertisers can tell their story through images, animation, text, music, video, and more, and promote that content by sponsoring it to appear as an advertisement in the streams of Tumblr users. In this paper we present a framework that enabled one of the key targeted advertising components for Tumblr, specifically gender and interest targeting. We describe the main challenges involved in development of the framework, which include creating the ground truth for training gender prediction models, as well as mapping Tumblr content to an interest taxonomy. For purposes of inferring user interests we propose a novel semi-supervised neural language model for categorization of Tumblr content (i.e., post tags and post keywords). The model was trained on a large-scale data set consisting of 6.8 billion user posts, with very limited amount of categorized keywords, and was shown to have superior performance over the bag-of-words model. We successfully deployed gender and interest targeting capability in Yahoo production systems, delivering inference for users that cover more than 90% of daily activities at Tumblr. Online performance results indicate advantages of the proposed approach, where we observed 20% lift in user engagement with sponsored posts as compared to untargeted campaigns.

Mihajlo Grbovic, Vladan Radosavljevic, Nemanja Djuric et al.

In recent years online advertising has become increasingly ubiquitous and effective. Advertisements shown to visitors fund sites and apps that publish digital content, manage social networks, and operate e-mail services. Given such large variety of internet resources, determining an appropriate type of advertising for a given platform has become critical to financial success. Native advertisements, namely ads that are similar in look and feel to content, have had great success in news and social feeds. However, to date there has not been a winning formula for ads in e-mail clients. In this paper we describe a system that leverages user purchase history determined from e-mail receipts to deliver highly personalized product ads to Yahoo Mail users. We propose to use a novel neural language-based algorithm specifically tailored for delivering effective product recommendations, which was evaluated against baselines that included showing popular products and products predicted based on co-occurrence. We conducted rigorous offline testing using a large-scale product purchase data set, covering purchases of more than 29 million users from 172 e-commerce websites. Ads in the form of product recommendations were successfully tested on online traffic, where we observed a steady 9% lift in click-through rates over other ad formats in mail, as well as comparable lift in conversion rates. Following successful tests, the system was launched into production during the holiday season of 2014.