Optimizing Etsy Advertisements with Live User Behavior for Personalized Experiences
Introduction:
Introduction
Personalization plays a crucial role in connecting our unique marketplace with the right buyers at the right time. Etsy has recently introduced an innovative approach to personalization through its adSformer Diversifiable Personalization Module (ADPM). This deep learning module encodes and learns from short-term sequences of user actions, allowing us to personalize clickthrough rates (CTR) and post-click conversion rates (PCCVR) ranking models used in Etsy Ads.
In this introduction, we provide an overview of Etsy’s personalized ranking system and how it enhances the relevance of sponsored listings. We analyze the challenges of session personalization and highlight the importance of utilizing user actions to understand their intent. The ADPM is introduced as a solution that leverages temporal and content signals to personalize user sessions effectively.
The ADPM consists of three components: the adSformer Encoder, pretrained representations, and representations learned “on the fly.” These components work together symbiotically to generate a dynamic representation of the user, incorporating recent actions and interactions within a session. The user representation is then employed in various recommenders and rankers for real-time personalization.
We also discuss the use of pretrained representations, including visual representations and Ads Information Retrieval (AIR) item representations. Visual representations capture rich image signals, while AIR item representations enable the learning of similarity metrics between items.
Overall, Etsy’s ADPM provides a powerful method for personalization, improving the relevance and effectiveness of sponsored listings on the platform. The combination of temporal and content signals enables Etsy to understand user intent and deliver personalized results, enhancing the shopping experience for both sellers and buyers.
Full Article: Optimizing Etsy Advertisements with Live User Behavior for Personalized Experiences
Etsy Implements Personalized ML Models for Better User Experience and Relevant Ads
Etsy, the online marketplace for handmade and vintage goods, has introduced a new method of personalizing machine learning (ML) models to improve the user experience and relevance of its ads. The company’s adSformer Diversifiable Personalization Module (ADPM) utilizes encoding and learning from short-term sequences of user actions to personalize clickthrough rates (CTR) and post-click conversion rates (PCCVR) ranking models.
Personalization in Etsy Ads
Etsy allows sellers to place sponsored listings alongside organic search results. To ensure the relevance of these ads, personalization is crucial. For example, if a user has recently interacted with men’s leather jackets, they would expect to see ads related to that specific interest, rather than generic jacket ads.
Understanding User Sessions
To personalize ads effectively, Etsy takes into account the concept of user sessions, defined as one-hour shopping windows. Analyzing a sample of sessions, Etsy found a power law distribution where most users interact with only a small number of listings during their session.
Leveraging User Actions
To overcome the challenge of limited listing views, Etsy leverages the rich stream of user actions that indicate intent, such as search queries, item favorites, views, add-to-carts, and purchases. These actions provide valuable insights into the user’s preferences and current interests.
The Role of Sequences
By analyzing the sequences of user actions, Etsy gains a deeper understanding of the user’s intent. For example, if a buyer has viewed a sequence of lamps, such as a 70s orange lamp, a retro table lamp, and a vintage mushroom lamp, Etsy can infer that the user is interested in lamps with specific characteristics and preferences.
The adSformer Diversifiable Personalization Module (ADPM)
ADPM is Etsy’s solution for session personalization using temporal and content signals. It consists of three components that work together to improve the relevancy of ads:
1. The adSformer Encoder: This component learns a deep representation of the user’s one-hour input sequence. It modifies the standard transformer block by adding a global max pooling layer to extract the most salient signals from the sequence.
2. Pretrained Representations: Etsy utilizes pretrained embeddings of item IDs, including image, text, and multimodal representations, to encode the user’s one-hour sequence of actions. These rich representations are learned offline and provide efficient and diverse information about the items.
3. Representations Learned “On the Fly”: This component learns lightweight representations for sequences that don’t have pretrained representations available. For example, Etsy learns embeddings for favorited shop IDs to capture user preferences.
Benefits and Implementation
ADPM’s three components work symbiotically to improve the relevancy of personalized results and enhance the user experience. The module is implemented as a Tensorflow Keras module, making it easily adaptable for downstream models. Etsy also leverages pretrained representations, such as image embeddings trained using multitask classification learning.
Improving Image Representations
Etsy employs image signals in various tasks, including visually similar candidate generation and search by image. To enhance the quality of image representations, the company is updating the backbone architectures to efficient vision transformers.
Ads Information Retrieval Representations
Etsy utilizes the Ads Information Retrieval (AIR) approach to encode item IDs through a metric learning approach. These representations capture the similarity between items, enhancing the relevancy of personalized ads.
Conclusion
Etsy’s implementation of personalized ML models through the adSformer Diversifiable Personalization Module (ADPM) is aimed at improving user experience and delivering relevant ads. By leveraging user actions, sequence analysis, and pretrained representations, Etsy ensures that its ads align with the user’s intent and preferences. The use of diversified signals and lightweight representations enhances the personalized results and contributes to the overall success of Etsy’s platform.
Summary: Optimizing Etsy Advertisements with Live User Behavior for Personalized Experiences
Summary: Etsy has introduced a new method for personalizing ML models called the adSformer Diversifiable Personalization Module (ADPM). This module aims to personalize the clickthrough rate (CTR) and post-click conversion rate (PCCVR) ranking models used in Etsy Ads by encoding and learning from short-term user actions. The module consists of three components: the adSformer Encoder, pretrained representations, and representations learned “on the fly.” These components work together to create a dynamic user representation that can be used to personalize the shopping experience in real-time. Etsy leverages rich streams of user actions, such as search queries, item favorites, and purchases, to understand the user’s interests and deliver relevant personalized results. The ADPM module is implemented as a Tensorflow Keras module and can be easily incorporated into downstream models. Etsy also utilizes pretrained representations, including image embeddings, text embeddings, and multimodal item representations, to enhance the quality of the user representation. The company is continually improving its visual representations by updating its backbone architectures to efficient vision transformers. Additionally, Etsy uses Ads Information Retrieval (AIR) item representations to learn a similarity metric between items, allowing for more interpretable results. Overall, the ADPM module and pretrained representations help Etsy personalize its marketplace and deliver relevant sponsored listings to users based on their intent.
Frequently Asked Questions:
Q1: What is machine learning and why is it important?
A: Machine learning is a subset of artificial intelligence that involves the design and development of algorithms and models capable of learning from data and making predictions or decisions without being explicitly programmed. It is important because it enables computers to automatically analyze and interpret vast amounts of data, extract meaningful patterns, and make accurate predictions or decisions.
Q2: How does machine learning work?
A: Machine learning algorithms use statistical techniques to identify patterns and relationships within datasets. They learn from the training data by adjusting their internal parameters until they can accurately predict or classify new data. These algorithms can be categorized as supervised, unsupervised, or reinforcement learning, depending on the type of available data and the learning approach used.
Q3: What are some practical applications of machine learning?
A: Machine learning finds applications in various domains, such as finance, healthcare, marketing, and self-driving cars. Some common examples include spam filtering, fraud detection, recommendation systems, image recognition, natural language processing, and predictive maintenance. It has the potential to revolutionize industries by enhancing decision-making, automating processes, and uncovering valuable insights from data.
Q4: What are the main challenges faced in machine learning?
A: Machine learning faces challenges such as overfitting (when a model performs well on training data but poorly on new data), selection bias (when the training data is not representative of the real-world population), and computational complexity (when handling large datasets or complex models). Additionally, ensuring privacy, fairness, and interpretability in machine learning models are important considerations.
Q5: How can machine learning models be evaluated?
A: Machine learning models are evaluated using various metrics depending on the task. For classification problems, metrics like accuracy, precision, recall, and F1 score are commonly used. For regression problems, metrics such as mean squared error (MSE) or root mean squared error (RMSE) are often employed. Cross-validation, where the dataset is split into multiple train-test sets, is a common technique to evaluate model performance and detect potential overfitting. Additionally, domain-specific evaluation metrics may exist for specific applications.
