Etsy Engineering | Deep Learning for Search Ranking at Etsy

Etsy Engineering Unveils Innovative Deep Learning for Enhanced Search Ranking on Etsy Platform

Introduction:

Introduction:

The search ranking at Etsy has undergone a significant transformation with the implementation of a deep learning model. Previously, the search ranking was powered by a gradient-boosted decision tree model, which served relevant results but showed diminishing returns over time. The decision to migrate to a neural ranking model brought about various challenges, including the need to modernize the development pipeline and infrastructure. Through iteration and experimentation, Etsy successfully launched its first-ever unified deep learning model for search ranking, providing decreased model training time and an improved developer experience. This post explores the journey taken to evolve the ranking model and the future direction of Etsy’s search ranking.

Full Article: Etsy Engineering Unveils Innovative Deep Learning for Enhanced Search Ranking on Etsy Platform

Etsy, an online marketplace known for its handmade and vintage items, recently made a significant change to its search ranking system. Previously, Etsy used a gradient-boosted decision tree model to power its search results. This model performed well and delivered personalized search results. However, over time, the model’s performance began to plateau, and the company saw diminishing returns.

To overcome this limitation, Etsy decided to migrate from the decision tree model to a deep learning model for search ranking. Deep learning models provide more flexibility and exploration for improvements beyond the constraints of traditional tree feature engineering. This migration presented significant challenges as it required major changes in the development pipelines and serving infrastructure.

The journey to implementing Etsy’s first-ever unified deep learning model for search ranking took exactly one year. During this time, the company iterated, experimented, and made adjustments until they finally launched the new model. They modernized their machine learning development pipeline and adopted open-source tools and functions, with TensorFlow’s learning-to-rank library at the core.

You May Also Like to Read  Unleashing the Power of ChatGPT: A Step-by-Step Guide in Alpaca Style for Training Your Own Chatbot - Part 1

Migrating to a neural ranking model brought several benefits for Etsy. The model training time decreased significantly, resulting in a faster and more efficient developer experience. Additionally, the deep learning model unlocked new types of features that were previously unavailable in the tree model.

Initially, Etsy considered using an ensemble model that combined the gradient-boosted decision tree model with a neural ranking model. The idea was to leverage the accumulated knowledge and engineered features of the tree model while allowing the neural network to learn additional information. However, the ensemble model did not show improvement over the baseline when tested offline. Moreover, implementing support for multiple models in parallel posed a significant development challenge and increased complexity.

After learning from the complications of the ensemble model, Etsy decided to focus on the capabilities of the neural network alone. Compared to an ensemble model, a single deep learning model had several advantages. It created a more unified and productive experience for developers, reduced model training time, and improved user experience by minimizing inference latency. Additionally, maintaining one model was less costly and added stability to the system.

To build the neural ranking model, Etsy incorporated the engineered features from the decision tree, such as click rate and listing price. They also added custom embeddings trained on Etsy data to represent text features. These embeddings proved to be more effective than off-the-shelf models for ranking results, highlighting the power of domain-specific trained embeddings.

The neural ranking model required longer training windows than the tree model but achieved better performance with more data. Etsy surpassed parity offline in their prototype of the new model, but replicating that success online presented a greater challenge. Serving the neural ranking model at production scale required careful planning and infrastructure adjustments.

You May Also Like to Read  Join the Apple Machine Learning for Health Workshop 2023 - Learn the Latest in Health Tech!

Overall, migrating from a decision tree model to a deep learning model for search ranking marked a significant milestone for Etsy. The new model offered improved performance, decreased model training time, and a more unified development experience. Etsy continues to explore and refine their ranking model as they strive to provide the best search results for their users.

Summary: Etsy Engineering Unveils Innovative Deep Learning for Enhanced Search Ranking on Etsy Platform

Summary:

Etsy recently migrated from a gradient-boosted decision tree model to a neural ranking model for search ranking. They found that decision trees were limited in their ability to provide relevant results, and deep learning models offered more potential for improvement. Initially, Etsy considered using an ensemble model with both the tree and neural models, but it proved to be complicated and time-consuming. They decided to focus on the neural model alone, which provided advantages such as faster training time and a unified development experience. They incorporated features from the decision tree model and experimented with text embeddings to improve relevancy. Offline testing showed promising results, and the next challenge is scaling the neural ranking model for production.

Frequently Asked Questions:

Q1: What is machine learning?
A1: Machine learning is a subset of artificial intelligence that focuses on enabling computer systems to learn and make decisions without being explicitly programmed. It involves developing algorithms and statistical models that allow machines to analyze and interpret data, learn from patterns, and improve their performance over time.

Q2: How does machine learning work?
A2: Machine learning algorithms typically rely on a training process where they are exposed to a large amount of data, with desired outcomes or labels associated with that data. The algorithm learns from this labeled data to recognize patterns and make predictions or decisions. As more data is processed, the machine learning model becomes more accurate and capable of performing tasks autonomously.

You May Also Like to Read  Speeding up our A/B experiments using machine learning: Enhancing efficiency and results

Q3: What are some real-life applications of machine learning?
A3: Machine learning has various applications across different industries. Some examples include:

1. Recommendation systems in e-commerce platforms, suggesting products based on user preferences.
2. Fraud detection systems in banking, flagging suspicious transactions based on historical patterns.
3. Speech recognition technologies used in virtual assistants like Siri or Alexa.
4. Computer vision algorithms used in autonomous vehicles or facial recognition systems.
5. Medical diagnosis systems that assist doctors in predicting diseases based on patient data.

Q4: What are the different types of machine learning algorithms?
A4: Machine learning algorithms can be broadly categorized into three main types:

1. Supervised learning: In this type, the algorithm learns from labeled data, making predictions based on known outcomes.
2. Unsupervised learning: Here, the algorithm explores unlabeled data to identify patterns or group similar data points together.
3. Reinforcement learning: In reinforcement learning, an agent learns to make decisions in an environment by receiving feedback or rewards for its actions.

Q5: What are the challenges in implementing machine learning?
A5: Implementing machine learning comes with its own set of challenges:

1. Data quality and quantity: Machine learning models heavily rely on large amounts of high-quality data. Gathering and preprocessing relevant data can be time-consuming and resource-intensive.
2. Model interpretability: Some machine learning algorithms, like deep neural networks, can be complex and difficult to interpret. Understanding how a model arrives at a decision can be a challenge, especially when it involves sensitive or critical applications.
3. Overfitting and generalization: Overfitting occurs when a model becomes too specific to the training data and fails to generalize well to unseen data. Striking a balance between model complexity and generalization is crucial.
4. Ethical considerations: Machine learning algorithms may inadvertently learn and perpetuate biases present in the data they were trained on. Ensuring fairness and transparency in machine learning systems is an important ethical consideration.

Note: The answers provided here are for illustrative purposes only and may require more comprehensive explanations depending on the target audience’s knowledge level.