arXiv subjects

Srijan Saket

Publications and source records attributed to Srijan Saket.

4 recordsLinked to original sources

arXiv2025

Dimension Mask Layer: Optimizing Embedding Efficiency for Scalable ID-based Models

In modern recommendation systems and social media platforms like Meta, TikTok, and Instagram, large-scale ID-based features often require embedding tables that consume significant memory. Managing these embedding sizes can be challenging, leading to bulky models that are harder to deploy and maintain. In this paper, we introduce a method to automatically determine the optimal embedding size for ID features, significantly reducing the model size while maintaining performance. Our approach involves defining a custom Keras layer called the dimension mask layer, which sits directly after the embedding lookup. This layer trims the embedding vector by allowing only the first N dimensions to pass through. By doing this, we can reduce the input feature dimension by more than half with minimal or no loss in model performance metrics. This reduction helps cut down the memory footprint of the model and lowers the risk of overfitting due to multicollinearity. Through offline experiments on public datasets and an online A/B test on a real production dataset, we demonstrate that using a dimension mask layer can shrink the effective embedding dimension by 40-50\%, leading to substantial improvements in memory efficiency. This method provides a scalable solution for platforms dealing with a high volume of ID features, optimizing both resource usage and model performance.

cs.IR↗

arXiv2024

Real-time Event Joining in Practice With Kafka and Flink

Historically, machine learning training pipelines have predominantly relied on batch training models, retraining models every few hours. However, industrial practitioners have proved that real-time training can lead to a more adaptive and personalized user experience. The transition from batch to real-time is full of tradeoffs to get the benefits of accuracy and freshness while keeping the costs low and having a predictable, maintainable system. Our work characterizes migrating to a streaming pipeline for a machine learning model using Apache Kafka and Flink. We demonstrate how to transition from Google Pub/Sub to Kafka to handle incoming real-time events and leverage Flink for streaming joins using RocksDB and checkpointing. We also address challenges such as managing causal dependencies between events, balancing event time versus processing time, and ensuring exactly-once versus at-least-once delivery guarantees, among other issues. Furthermore, we showcase how we improved scalability by using topic partitioning in Kafka, reduced event throughput by \textbf{85\%} through the use of Avro schema and compression, decreased costs by \textbf{40\%}, and implemented a separate pipeline to ensure data correctness. Our findings provide valuable insights into the tradeoffs and complexities of real-time systems, enabling better-informed decisions tailored to specific requirements for building effective streaming systems that enhance user satisfaction.

cs.SE↗

arXiv2024

Crafting Tomorrow: The Influence of Design Choices on Fresh Content in Social Media Recommendation

The rise in popularity of social media platforms, has resulted in millions of new, content pieces being created every day. This surge in content creation underscores the need to pay attention to our design choices as they can greatly impact how long content remains relevant. In today's landscape where regularly recommending new content is crucial, particularly in the absence of detailed information, a variety of factors such as UI features, algorithms and system settings contribute to shaping the journey of content across the platform. While previous research has focused on how new content affects users' experiences, this study takes a different approach by analyzing these decisions considering the content itself. Through a series of carefully crafted experiments we explore how seemingly small decisions can influence the longevity of content, measured by metrics like Content Progression (CVP) and Content Survival (CSR). We also emphasize the importance of recognizing the stages that content goes through underscoring the need to tailor strategies for each stage as a one size fits all approach may not be effective. Additionally we argue for a departure from traditional experimental setups in the study of content lifecycles, to avoid potential misunderstandings while proposing advanced techniques, to achieve greater precision and accuracy in the evaluation process.

cs.IR↗

arXiv2024

Monitoring the Evolution of Behavioural Embeddings in Social Media Recommendation

Emerging short-video platforms like TikTok, Instagram Reels, and ShareChat present unique challenges for recommender systems, primarily originating from a continuous stream of new content. ShareChat alone receives approximately 2 million pieces of fresh content daily, complicating efforts to assess quality, learn effective latent representations, and accurately match content with the appropriate user base, especially given limited user feedback. Embedding-based approaches are a popular choice for industrial recommender systems because they can learn low-dimensional representations of items, leading to effective recommendation that can easily scale to millions of items and users. Our work characterizes the evolution of such embeddings in short-video recommendation systems, comparing the effect of batch and real-time updates to content embeddings. We investigate \emph{how} embeddings change with subsequent updates, explore the relationship between embeddings and popularity bias, and highlight their impact on user engagement metrics. Our study unveils the contrast in the number of interactions needed to achieve mature embeddings in a batch learning setup versus a real-time one, identifies the point of highest information updates, and explores the distribution of $\ell_2$-norms across the two competing learning modes. Utilizing a production system deployed on a large-scale short-video app with over 180 million users, our findings offer insights into designing effective recommendation systems and enhancing user satisfaction and engagement in short-video applications.

cs.IR↗