The internet offers a massive repository of unstructured information, but it's a significant challenge to convert this into a structured format. At Pinterest, the ability to accurately extract structured product data from e-commerce websites is essential to enhance user experiences and improve content distribution. In this paper, we present Pinterest's system for attribute extraction, which achieves remarkable accuracy and scalability at a manageable cost. Our approach leverages a novel webpage representation that combines structural, visual, and text modalities into a compact form, optimizing it for small model learning. This representation captures each visible HTML node with its text, style and layout information. We show how this allows simple models such as eXtreme Gradient Boosting (XGBoost) to extract attributes more accurately than much more complex Large Language Models (LLMs) such as Generative Pre-trained Transformer (GPT). Our results demonstrate a system that is highly scalable, processing over 1,000 URLs per second, while being 1000 times more cost-effective than the cheapest GPT alternatives.

The ergodic interference alignment scheme allows each user in a K-user interference channel to attain half its interference-free rate at any signal-to-noise ratio (SNR). However, this comes at the cost of extraordinarily long delays, since each channel matrix must be paired with a complementary matrix that can cancel out its off-diagonal terms. In this paper, we investigate the opposite approach. Rather than waiting for perfect matches, we simply try to pair together the available channel matrices to attain the highest possible sum rate. We show that this problem is closely connected to the problem of maximum weight matching, which allows us to efficiently evaluate the numerical performance of our scheme.

Consider a three-user time-varying Gaussian interference channel under uniform phase fading. Prior work on ergodic interference alignment has shown that each user can simultaneously achieve half its interference-free rate. In this paper, we combine ideas from ergodic alignment and compute-and-forward to characterize the symmetric ergodic capacity of this channel in the strong and very strong regimes to within a constant gap.