Large Language Models (LLMs) pose a new paradigm of modeling and computation for information tasks. Recommendation systems are a critical application domain poised to benefit significantly from the sequence modeling capabilities and world knowledge inherent in these large models. In this paper, we introduce PLUM, a framework designed to adapt pre-trained LLMs for industry-scale recommendation tasks. PLUM consists of item tokenization using Semantic IDs, continued pre-training (CPT) on domain-specific data, and task-specific fine-tuning for recommendation objectives. For fine-tuning, we focus particularly on generative retrieval, where the model is directly trained to generate Semantic IDs of recommended items based on user context. We conduct comprehensive experiments on large-scale internal video recommendation datasets. Our results demonstrate that PLUM achieves substantial improvements for retrieval compared to a heavily-optimized production model built with large embedding tables. We also present a scaling study for the model's retrieval performance, our learnings about CPT, a few enhancements to Semantic IDs, along with an overview of the training and inference methods that enable launching this framework to billions of users in YouTube.

Sequential models are invaluable for powering personalized recommendation systems. In the context of short-form video (SFV) feeds, where user behavior history is typically longer, systems must be able to understand users’ long-term interests. However, deploying large sequence models to extensive web-scale applications faces challenges due to high serving cost. To address this, we propose an industrial framework designed for efficiently serving large user sequence models. Specifically, the proposed infrastructure decouples serving of the user sequence model and the main recommendation model, with the user sequence model being served offline (asynchronously) with periodical refresh. The proposed infrastructure is also model-agnostic; thus, it can be used to support any type of user sequence models (even LLMs) with controllable costs. Empirical results show that large user models deployed with our framework significantly and consistently enhance the quality of the main recommendation model with minimal serving costs increase.

Randomly-hashed item ids are used ubiquitously in recommendation models. However, the learned representations from random hashing prevents generalization across similar items, causing problems of learning unseen and long-tail items, especially when item corpus is large, power-law distributed, and evolving dynamically. In this paper, we propose using content-derived features as a replacement for random ids. We show that simply replacing ID features with content-based embeddings can cause a drop in quality due to reduced memorization capability. To strike a good balance of memorization and generalization, we propose to use Semantic IDs [15], a compact and discrete item representation, as a replacement for random item ids. Semantic IDs are learned from frozen content embeddings using RQ-VAE and thus can capture the hierarchy of concepts in items. Similar to content embeddings, the compactness of Semantic IDs poses a problem of adaption in recommendation models. We propose novel methods for adapting Semantic IDs in industry-scale ranking models, through hashing sub-pieces of of the Semantic-ID sequences. In particular, we find that the SentencePiece model [10] that is commonly used in LLM tokenization outperforms manually crafted pieces such as N-grams. To the end, we evaluate our approaches in a real-world ranking model for YouTube recommendations. Our experiments demonstrate that Semantic IDs can replace the direct use of video IDs by improving the generalization ability on new and long-tail item slices without sacrificing overall model quality.

Industry recommender systems usually suffer from highly-skewed long-tail item distributions where a small fraction of the items receives most of the user feedback. This skew hurts recommender quality especially for the item slices without much user feedback. While there have been many research advances made in academia, deploying these methods in production is very difficult and very few improvements have been made in industry. One challenge is that these methods often hurt overall performance; additionally, they could be complex and expensive to train and serve.

Recommender systems play an important role in many content platforms. While most recommendation research is dedicated to designing better models to improve user experience, we found that research on stabilizing the training for such models is severely under-explored. As recommendation models become larger and more sophisticated, they are more susceptible to training instability issues, i.e., loss divergence, which can make the model unusable, waste significant resources and block model developments. In this paper, we share our findings and best practices we learned for improving the training stability of a real-world multitask ranking model for YouTube recommendations. We show some properties of the model that lead to unstable training and conjecture on the causes. Furthermore, based on our observations of training dynamics near the point of training instability, we hypothesize why existing solutions would fail, and propose a new algorithm to mitigate the limitations of existing solutions. Our experiments on YouTube production dataset show the proposed algorithm can significantly improve training stability while not compromising convergence, comparing with several commonly used baseline methods.

The last decade has witnessed many successes of deep learning-based models for industry-scale recommender systems. These models are typically trained offline in a batch manner. While being effective in capturing users’ past interactions with recommendation platforms, batch learning suffers from long model-update latency and is vulnerable to system biases, making it hard to adapt to distribution shift and explore new items or user interests. Although online learning-based approaches (e.g., multi-armed bandits) have demonstrated promising theoretical results in tackling these challenges, their practical real-time implementation in large-scale recommender systems remains limited. First, the scalability of online approaches in servicing a massive online traffic while ensuring timely updates of bandit parameters poses a significant challenge. Additionally, exploring uncertainty in recommender systems can easily result in unfavorable user experience, highlighting the need for devising intricate strategies that effectively balance the trade-off between exploitation and exploration. In this paper, we introduce Online Matching: a scalable closed-loop bandit system learning from users’ direct feedback on items in real time. We present a hybrid offline + online approach for constructing this system, accompanied by a comprehensive exposition of the end-to-end system architecture. We propose Diag-LinUCB – a novel extension of the LinUCB algorithm – to enable distributed updates of bandits parameter in a scalable and timely manner. We conduct live experiments in YouTube and show that Online Matching is able to enhance the capabilities of fresh content discovery and item exploration in the present platform.

Multi-Task Learning (MTL) is a powerful learning paradigm to improve generalization performance via knowledge sharing. However, existing studies find that MTL could sometimes hurt generalization, especially when two tasks are less correlated. One possible reason that hurts generalization is spurious correlation, i.e., some knowledge is spurious and not causally related to task labels, but the model could mistakenly utilize them and thus fail when such correlation changes. In MTL setup, there exist several unique challenges of spurious correlation. First, the risk of having non-causal knowledge is higher, as the shared MTL model needs to encode all knowledge from different tasks, and causal knowledge for one task could be potentially spurious to the other. Second, the confounder between task labels brings in a different type of spurious correlation to MTL. Given such label-label confounders, we theoretically and empirically show that MTL is prone to taking non-causal knowledge from other tasks. To solve this problem, we propose Multi-Task Causal Representation Learning (MT-CRL) framework. MT-CRL aims to represent multi-task knowledge via disentangled neural modules, and learn which module is causally related to each task via MTL-specific invariant regularization. Experiments show that MT-CRL could enhance MTL model's performance by 5.5% on average over Multi-MNIST, MovieLens, Taskonomy, CityScape, and NYUv2, and show it could indeed alleviate spurious correlation problem.

Modern recommender systems have evolved rapidly along with deep learning models that are well-optimized for overall performance, especially those trained under Empirical Risk Minimization (ERM). However, a recommendation algorithm that focuses solely on the average performance may reinforce the exposure bias and exacerbate the “rich-get-richer” effect, leading to unfair user experience. In a simulation study, we demonstrate that such performance gap among various user groups is enlarged by an ERM-trained recommender in the long-term. To mitigate such amplification effects, we propose to optimize for the worst-case performance under the Distributionally Robust Optimization (DRO) framework, with the goal of improving long-term fairness for disadvantaged subgroups. In addition, we propose a simple-yet-effective streaming optimization improvement called Streaming-DRO (S-DRO), which effectively reduces loss variances for recommendation problems with sparse and long-tailed data distributions. Our results on two large-scale datasets suggest that (1) DRO is a flexible and effective technique for improving worst-case performance, and (2) Streaming-DRO outperforms vanilla DRO and other strong baselines by improving the worst-case and overall performance at the same time.

We consider the problem of listwise learning-to-rank (LTR) on data with \textit{partitioned preference}, where a set of items are sliced into ordered and disjoint partitions, but the ranking of items within a partition is unknown. The Plackett-Luce (PL) model has been widely used in listwise LTR methods. However, given $N$ items with $M$ partitions, calculating the likelihood of data with partitioned preference under the PL model has a time complexity of $O(N+S!)$, where $S$ is the maximum size of the top $M-1$ partitions. This computational challenge restrains existing PL-based listwise LTR methods to only a special case of partitioned preference, \textit{top-$K$ ranking}, where the exact order of the top $K$ items is known. In this paper, we exploit a random utility model formulation of the PL model and propose an efficient approach through numerical integration for calculating the likelihood. This numerical approach reduces the aforementioned time complexity to $O(N+MS)$, which allows training deep-neural-network-based ranking models with a large output space. We demonstrate that the proposed method outperforms well-known LTR baselines and remains scalable through both simulation experiments and applications to real-world eXtreme Multi-Label (XML) classification tasks. The proposed method also achieves state-of-the-art performance on XML datasets with relatively large numbers of labels per sample.

Large scale recommender models find most relevant items from huge catalogs, and they play a critical role in modern search and recommendation systems. To model the input space with large-vocab categorical features, a typical recommender model learns a joint embedding space through neural networks for both queries and items from user feedback data. However, with millions to billions of items in the corpus, users tend to provide feedback for a very small set of them, causing a power-law distribution. This makes the feedback data for long-tail items extremely sparse.

Embedding learning of categorical features (e.g. user/item IDs) is at the core of various recommendation models. The standard approach creates an embedding table where each row represents a dedicated embedding vector for every unique feature value. However, this method fails to efficiently handle high-cardinality features and unseen feature values (e.g. new video ID) that are prevalent in real-world recommendation systems. In this paper, we propose an alternative embedding framework Deep Hash Embedding (DHE), replacing embedding tables by a deep embedding network to compute embeddings on the fly. DHE first encodes the feature value to a unique identifier vector with multiple hashing functions and transformations, and then applies a DNN to convert the identifier vector to an embedding. The encoding module is deterministic, non-learnable, and free of storage, while the embedding network is updated during the training time to learn embedding generation. Empirical results show that DHE achieves comparable AUC against the standard one-hot full embedding, with smaller model sizes. Our work sheds light on the design of DNN-based alternative embedding schemes for categorical features without using embedding table lookup.

Highly skewed long-tail item distribution is very common in recommendation systems. It significantly hurts model performance on tail items. To improve tail-item recommendation, we conduct research to transfer knowledge from head items to tail items, leveraging the rich user feedback in head items and the semantic connections between head and tail items. Specifically, we propose a novel dual transfer learning framework that jointly learns the knowledge transfer from both model-level and item-level: 1. The model-level knowledge transfer builds a generic meta-mapping of model parameters from few-shot to many-shot model. It captures the implicit data augmentation on the model-level to improve the representation learning of tail items. 2. The item-level transfer connects head and tail items through item-level features, to ensure a smooth transfer of meta-mapping from head items to tail items. The two types of transfers are incorporated to ensure the learned knowledge from head items can be well applied for tail item representation learning in the long-tail distribution settings. Through extensive experiments on two benchmark datasets, results show that our proposed dual transfer learning framework significantly outperforms other state-of-the-art methods for tail item recommendation in hit ratio and NDCG. It is also very encouraging that our framework further improves head items and overall performance on top of the gains on tail items.

Modern online content-sharing platforms host billions of items like music, videos, and products uploaded by various providers for users to discover items of their interests. To satisfy the information needs, the task of effective item retrieval (or item search ranking) given user search queries has become one of the most fundamental problems to online content-sharing platforms. Moreover, the same query can represent different search intents for different users, so personalization is also essential for providing more satisfactory search results. Different from other similar research tasks, such as ad-hoc retrieval and product retrieval with copious words and reviews, items in content-sharing platforms usually lack sufficient descriptive information and related meta-data as features. In this paper, we propose the end-to-end deep attentive model (EDAM) to deal with personalized item retrieval for online content-sharing platforms using only discrete personal item history and queries. Each discrete item in the personal item history of a user and its content provider are first mapped to embedding vectors as continuous representations. A query-aware attention mechanism is then applied to identify the relevant contexts in the user history and construct the overall personal representation for a given query. Finally, an extreme multi-class softmax classifier aggregates the representations of both query and personal item history to provide personalized search results. We conduct extensive experiments on a large-scale real-world dataset with hundreds of million users from a large video media platform at Google. The experimental results demonstrate that our proposed approach significantly outperforms several competitive baseline methods. It is also worth mentioning that this work utilizes a massive dataset from a real-world commercial content-sharing platform for personalized item retrieval to provide more insightful analysis from the industrial aspects.

Many real-world recommender systems need to be highly scalable: matching millions of items with billions of users, with milliseconds latency. The scalability requirement has led to widely used two-stage recommender systems, consisting of efficient candidate generation model(s) in the first stage and a more powerful ranking model in the second stage.

Many recommendation systems retrieve and score items from a very large corpus. A common recipe to handle data sparsity and power-law item distribution is to learn item representations from its content features. Apart from many content-aware systems based on matrix factorization, we consider a modeling framework using two-tower neural net, with one of the towers (item tower) encoding a wide variety of item content features. A general recipe of training such two-tower models is to optimize loss functions calculated from in-batch negatives, which are items sampled from a random mini-batch. However, in-batch loss is subject to sampling biases, potentially hurting model performance, particularly in the case of highly skewed distribution. In this paper, we present a novel algorithm for estimating item frequency from streaming data. Through theoretical analysis and simulation, we show that the proposed algorithm can work without requiring fixed item vocabulary, and is capable of producing unbiased estimation and being adaptive to item distribution change. We then apply the sampling-bias-corrected modeling approach to build a large scale neural retrieval system for YouTube recommendations. The system is deployed to retrieve personalized suggestions from a corpus with tens of millions of videos. We demonstrate the effectiveness of sampling-bias correction through offline experiments on two real-world datasets. We also conduct live A/B testings to show that the neural retrieval system leads to improved recommendation quality for YouTube.

In this paper, we introduce a large scale multi-objective ranking system for recommending what video to watch next on an industrial video sharing platform. The system faces many real-world challenges, including the presence of multiple competing ranking objectives, as well as implicit selection biases in user feedback. To tackle these challenges, we explored a variety of soft-parameter sharing techniques such as Multi-gate Mixture-of-Experts so as to efficiently optimize for multiple ranking objectives. Additionally, we mitigated the selection biases by adopting a Wide & Deep framework. We demonstrated that our proposed techniques can lead to substantial improvements on recommendation quality on one of the world's largest video sharing platforms.

Neural-based multi-task learning has been successfully used in many real-world large-scale applications such as recommendation systems. For example, in movie recommendations, beyond providing users movies which they tend to purchase and watch, the system might also optimize for users liking the movies afterwards. With multi-task learning, we aim to build a single model that learns these multiple goals and tasks simultaneously. However, the prediction quality of commonly used multi-task models is often sensitive to the relationships between tasks. It is therefore important to study the modeling tradeoffs between task-specific objectives and inter-task relationships. In this work, we propose a novel multi-task learning approach, Multi-gate Mixture-of-Experts (MMoE), which explicitly learns to model task relationships from data. We adapt the Mixture-of-Experts (MoE) structure to multi-task learning by sharing the expert submodels across all tasks, while also having a gating network trained to optimize each task. To validate our approach on data with different levels of task relatedness, we first apply it to a synthetic dataset where we control the task relatedness. We show that the proposed approach performs better than baseline methods when the tasks are less related. We also show that the MMoE structure results in an additional trainability benefit, depending on different levels of randomness in the training data and model initialization. Furthermore, we demonstrate the performance improvements by MMoE on real tasks including a binary classification benchmark, and a large-scale content recommendation system at Google.

We consider the mixed regression problem with two components, under adversarial and stochastic noise. We give a convex optimization formulation that provably recovers the true solution, as well as a nonconvex formulation that works under more general settings and remains tractable. Upper bounds are provided on the recovery errors for both arbitrary noise and stochastic noise models. We also give matching minimax lower bounds (up to log factors), showing that our algorithm is information-theoretical optimal in a precise sense. Our results represent the first tractable algorithm guaranteeing successful recovery with tight bounds on recovery errors and sample complexity. Moreover, we pinpoint the statistical cost of mixtures: our minimax-optimal results indicate that the mixture poses a fundamentally more difficult problem in the low-SNR regime, where the learning rate changes.

We present minimax bounds for classification and clustering error in the setting where covariates are drawn from a mixture of two isotropic Gaussian distributions. Here, we define clustering error in a discriminative fashion, demonstrating fundamental connections between classification (supervised) and clustering (unsupervised). For both classification and clustering, our lower bounds show that without enough samples, the best any classifier or clustering rule can do is close to random guessing. For classification, as part of our upper bound analysis, we show that Fisher's linear discriminant achieves a fast minimax rate $\Theta(1/n)$ with enough samples $n$. For clustering, as part of our upper bound analysis, we show that a clustering rule constructed using principal component analysis achieves the minimax rate with enough samples. We also provide lower and upper bounds for the high-dimensional sparse setting where the dimensionality of the covariates $p$ is potentially larger than the number of samples $n$, but where the difference between the Gaussian means is sparse.

We consider the problem of Robust PCA in the fully and partially observed settings. Without corruptions, this is the well-known matrix completion problem. From a statistical standpoint this problem has been recently well-studied, and conditions on when recovery is possible (how many observations do we need, how many corruptions can we tolerate) via polynomial-time algorithms is by now understood. This paper presents and analyzes a non-convex optimization approach that greatly reduces the computational complexity of the above problems, compared to the best available algorithms. In particular, in the fully observed case, with $r$ denoting rank and $d$ dimension, we reduce the complexity from $O(r^2d^2\log(1/\epsilon))$ to $O(rd^2\log(1/\epsilon))$ -- a big savings when the rank is big. For the partially observed case, we show the complexity of our algorithm is no more than $O(r^4d\log(d)\log(1/\epsilon))$. Not only is this the best-known run-time for a provable algorithm under partial observation, but in the setting where $r$ is small compared to $d$, it also allows for near-linear-in-$d$ run-time that can be exploited in the fully-observed case as well, by simply running our algorithm on a subset of the observations.

We consider the weakly supervised binary classification problem where the labels are randomly flipped with probability $1-\alpha$. Although there exist numerous algorithms for this problem, it remains theoretically unexplored how the statistical accuracies and computational efficiency of these algorithms depend on the degree of supervision, which is quantified by $\alpha$. In this paper, we characterize the effect of $\alpha$ by establishing the information-theoretic and computational boundaries, namely, the minimax-optimal statistical accuracy that can be achieved by all algorithms, and polynomial-time algorithms under an oracle computational model. For small $\alpha$, our result shows a gap between these two boundaries, which represents the computational price of achieving the information-theoretic boundary due to the lack of supervision. Interestingly, we also show that this gap narrows as $\alpha$ increases. In other words, having more supervision, i.e., more correct labels, not only improves the optimal statistical accuracy as expected, but also enhances the computational efficiency for achieving such accuracy.

Model order reduction exploiting the spectral properties of the admittance matrix, known as the graph Laplacian, to control the approximation accuracy is a promising new class of approaches to power grid analysis. In this paper we introduce a method that allows a dramatic increase in the resulting graph sparsity and can handle large dense input graphs. The method is based on the observation that the information about the realistic ranges of port currents can be used to significantly improve the resulting graph sparsity. In practice, port currents cannot vary unboundedly and the estimates of peak currents are often available early in the design cycle. However, the existing methods including the sampling-based spectral sparsification approach [11] cannot utilize this information.

Binary embedding is a nonlinear dimension reduction methodology where high dimensional data are embedded into the Hamming cube while preserving the structure of the original space. Specifically, for an arbitrary N distinct points in \mathbbS^p-1, our goal is to encode each point using m-dimensional binary strings such that we can reconstruct their geodesic distance up to δuniform distortion. Existing binary embedding algorithms either lack theoretical guarantees or suffer from running time O(mp). We make three contributions: (1) we establish a lower bound that shows any binary embedding oblivious to the set of points requires m =Ω(\frac1δ^2\logN) bits and a similar lower bound for non-oblivious embeddings into Hamming distance; (2) we propose a novel fast binary embedding algorithm with provably optimal bit complexity m = O(\frac1 δ^2\logN) and near linear running time O(p \log p) whenever \log N ≪δ\sqrtp, with a slightly worse running time for larger \log N; (3) we also provide an analytic result about embedding a general set of points K ⊆\mathbbS^p-1 with even infinite size. Our theoretical findings are supported through experiments on both synthetic and real data sets.

Latent models are a fundamental modeling tool in machine learning applications, but they present significant computational and analytical challenges. The popular EM algorithm and its variants, is a much used algorithmic tool; yet our rigorous understanding of its performance is highly incomplete. Recently, work in [1] has demonstrated that for an important class of problems, EM exhibits linear local convergence. In the high-dimensional setting, however, the M-step may not be well defined. We address precisely this setting through a unified treatment using regularization. While regularization for high-dimensional problems is by now well understood, the iterative EM algorithm requires a careful balancing of making progress towards the solution while identifying the right structure (e.g., sparsity or low-rank). In particular, regularizing the M-step using the state-of-the-art high-dimensional prescriptions (e.g., `a la [19]) is not guaranteed to provide this balance. Our algorithm and analysis are linked in a way that reveals the balance between optimization and statistical errors. We specialize our general framework to sparse gaussian mixture models, high-dimensional mixed regression, and regression with missing variables, obtaining statistical guarantees for each of these examples.

Linear regression studies the problem of estimating a model parameter $\beta^* \in \R^p$, from $n$ observations $\{(y_i,x_i)\}_{i=1}^n$ from linear model $y_i = \langle \x_i,\beta^* \rangle + \epsilon_i$. We consider a significant generalization in which the relationship between $\langle x_i,\beta^* \rangle$ and $y_i$ is noisy, quantized to a single bit, potentially nonlinear, noninvertible, as well as unknown. This model is known as the single-index model in statistics, and, among other things, it represents a significant generalization of one-bit compressed sensing. We propose a novel spectral-based estimation procedure and show that we can recover $\beta^*$ in settings (i.e., classes of link function $f$) where previous algorithms fail. In general, our algorithm requires only very mild restrictions on the (unknown) functional relationship between $y_i$ and $\langle x_i,\beta^* \rangle$. We also consider the high dimensional setting where $\beta^*$ is sparse, and introduce a two-stage nonconvex framework that addresses estimation challenges in high dimensional regimes where $p \gg n$. For a broad class of link functions between $\langle x_i,\beta^* \rangle$ and $y_i$, we establish minimax lower bounds that demonstrate the optimality of our estimators in both the classical and high dimensional regimes.

We consider the mixed regression problem with two components, under adversarial and stochastic noise. We give a convex optimization formulation that provably recovers the true solution, and provide upper bounds on the recovery errors for both arbitrary noise and stochastic noise settings. We also give matching minimax lower bounds (up to log factors), showing that under certain assumptions, our algorithm is information-theoretically optimal. Our results represent the first (and currently only known) tractable algorithm guaranteeing successful recovery with tight bounds on recovery errors and sample complexity.

Mixed linear regression involves the recovery of two (or more) unknown vectors from unlabeled linear measurements; that is, where each sample comes from exactly one of the vectors, but we do not know which one. It is a classic problem, and the natural and empirically most popular approach to its solution has been the EM algorithm. As in other settings, this is prone to bad local minima; however, each iteration is very fast (alternating between guessing labels, and solving with those labels). In this paper we provide a new initialization procedure for EM, based on finding the leading two eigenvectors of an appropriate matrix. We then show that with this, a re-sampled version of the EM algorithm provably converges to the correct vectors, under natural assumptions on the sampling distribution, and with nearly optimal (unimprovable) sample complexity. This provides not only the first characterization of EM’s performance, but also much lower sample complexity as compared to both standard (randomly initialized) EM, and other methods for this problem.