Hybrid Preference Optimization: Augmenting Direct Preference Optimization with Auxiliary Objectives

For aligning large language models (LLMs), prior work has leveraged reinforcement learning via human feedback (RLHF) or variations of direct preference optimization (DPO). While DPO offers a simpler framework based on maximum likelihood estimation, it compromises on the ability to tune language models to easily maximize non-differentiable and non-binary objectives according to the LLM designer's preferences (e.g., using simpler language or minimizing specific kinds of harmful content). These may neither align with user preferences nor even be able to be captured tractably by binary preference data. To leverage the simplicity and performance of DPO with the generalizability of RL, we propose a hybrid approach between DPO and RLHF. With a simple augmentation to the implicit reward decomposition of DPO, we allow for tuning LLMs to maximize a set of arbitrary auxiliary rewards using offline RL. The proposed method, Hybrid Preference Optimization (HPO), shows the ability to effectively generalize to both user preferences and auxiliary designer objectives, while preserving alignment performance across a range of challenging benchmarks and model sizes.

OmniSearchSage: Multi-Task Multi-Entity Embeddings for Pinterest Search

In this paper, we present OmniSearchSage, a versatile and scalable system for understanding search queries, pins, and products for Pinterest search. We jointly learn a unified query embedding coupled with pin and product embeddings, leading to an improvement of $>8\%$ relevance, $>7\%$ engagement, and $>5\%$ ads CTR in Pinterest's production search system. The main contributors to these gains are improved content understanding, better multi-task learning, and real-time serving. We enrich our entity representations using diverse text derived from image captions from a generative LLM, historical engagement, and user-curated boards. Our multitask learning setup produces a single search query embedding in the same space as pin and product embeddings and compatible with pre-existing pin and product embeddings. We show the value of each feature through ablation studies, and show the effectiveness of a unified model compared to standalone counterparts. Finally, we share how these embeddings have been deployed across the Pinterest search stack, from retrieval to ranking, scaling to serve $300k$ requests per second at low latency. Our implementation of this work is available at https://github.com/pinterest/atg-research/tree/main/omnisearchsage.

Exploring intra-task relations to improve meta-learning algorithms

Meta-learning has emerged as an effective methodology to model several real-world tasks and problems due to its extraordinary effectiveness in the low-data regime. There are many scenarios ranging from the classification of rare diseases to language modelling of uncommon languages where the availability of large datasets is rare. Similarly, for more broader scenarios like self-driving, an autonomous vehicle needs to be trained to handle every situation well. This requires training the ML model on a variety of tasks with good quality data. But often times, we find that the data distribution across various tasks is skewed, i.e.the data follows a long-tail distribution. This leads to the model performing well on some tasks and not performing so well on others leading to model robustness issues. Meta-learning has recently emerged as a potential learning paradigm which can effectively learn from one task and generalize that learning to unseen tasks. In this study, we aim to exploit external knowledge of task relations to improve training stability via effective mini-batching of tasks. We hypothesize that selecting a diverse set of tasks in a mini-batch will lead to a better estimate of the full gradient and hence will lead to a reduction of noise in training.

Detecting anxiety from short clips of free-form speech

Barriers to accessing mental health assessments including cost and stigma continues to be an impediment in mental health diagnosis and treatment. Machine learning approaches based on speech samples could help in this direction. In this work, we develop machine learning solutions to diagnose anxiety disorders from audio journals of patients. We work on a novel anxiety dataset (provided through collaboration with Kintsugi Mindful Wellness Inc.) and experiment with several models of varying complexity utilizing audio, text and a combination of multiple modalities. We show that the multi-modal and audio embeddings based approaches achieve good performance in the task achieving an AUC ROC score of 0.68-0.69.

Exploring Graph Based Approaches for Author Name Disambiguation

In many applications, such as scientific literature management, researcher search, social network analysis and etc, Name Disambiguation (aiming at disambiguating WhoIsWho) has been a challenging problem. In addition, the growth of scientific literature makes the problem more difficult and urgent. Although name disambiguation has been extensively studied in academia and industry, the problem has not been solved well due to the clutter of data and the complexity of the same name scenario. In this work, we aim to explore models that can perform the task of name disambiguation using the network structure that is intrinsic to the problem and present an analysis of the models.

Modeling User Behavior With Interaction Networks for Spam Detection

Spam is a serious problem plaguing web-scale digital platforms which facilitate user content creation and distribution. It compromises platform's integrity, performance of services like recommendation and search, and overall business. Spammers engage in a variety of abusive and evasive behavior which are distinct from non-spammers. Users' complex behavior can be well represented by a heterogeneous graph rich with node and edge attributes. Learning to identify spammers in such a graph for a web-scale platform is challenging because of its structural complexity and size. In this paper, we propose SEINE (Spam DEtection using Interaction NEtworks), a spam detection model over a novel graph framework. Our graph simultaneously captures rich users' details and behavior and enables learning on a billion-scale graph. Our model considers neighborhood along with edge types and attributes, allowing it to capture a wide range of spammers. SEINE, trained on a real dataset of tens of millions of nodes and billions of edges, achieves a high performance of 80% recall with 1% false positive rate. SEINE achieves comparable performance to the state-of-the-art techniques on a public dataset while being pragmatic to be used in a large-scale production system.