CTR Prediction Model Using User, Item, and Query Embeddings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional collaborative filtering systems for online concierge systems rely on binary matrices to predict user interests based on limited actions like clicking on ads, which do not effectively leverage multiple interaction signals for improving click-through rates in search ads.
Innovation Solution
An online concierge system uses a machine learning model that generates user and item embeddings, incorporating various interaction strengths such as purchasing, adding to a shopping cart, and clicking, to rank promoted items based on search queries, employing a Hadamard product and click-through rate model to select and display relevant items.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional collaborative filtering systems use binary matrices with single signals (clicking, adding to cart, or purchasing) to predict user interests, then the system implementation is simple, but the prediction accuracy and click-through rate for search ads are insufficient
Solution Approach 1:
The patent transforms the conventional binary matrix representation into a multi-dimensional embedding space where users, items, and interactions are represented as continuous vectors. This dimensional transformation allows the system to capture nuanced interaction strengths (e.g., purchase vs. click) and complex relationships that binary matrices cannot represent, thereby improving prediction accuracy while managing complexity through efficient embedding techniques.
Solution Approach 2:
The patent combines multiple interaction signals (clicking, adding to cart, purchasing) into a unified embedding representation that integrates different types of user behaviors. By composing these diverse signals into a composite embedding vector, the system leverages the strengths of each interaction type to improve overall prediction accuracy without simply adding more separate systems.
2Productivity
If the system uses multiple interaction signals with different strengths (purchasing > adding to cart > clicking) to improve ad relevance, then the click-through rate improves, but the computational complexity increases
Solution Approach 1:
The patent pre-computes and stores user embeddings, item embeddings, and interaction embeddings in advance before the actual ad serving process. By performing these computationally intensive embedding calculations beforehand and caching the results, the system reduces real-time computational complexity while maintaining the ability to leverage multiple interaction signals for improved click-through rates.
Solution Approach 2:
The patent replaces complex mechanical computation of interaction strengths with learned embedding representations that are computed through neural network models. Instead of explicitly calculating interaction strengths using traditional methods, the system uses embedding models to automatically learn and represent interaction patterns, reducing computational overhead while improving accuracy.
Data Source
AI summary
An online concierge system uses a machine learning click through rate model to select promoted items based on user embeddings, item embeddings, and search query embeddings. Embeddings obtained by an embedding model may be used as inputs to the click through rate model. The embedding model may be trained using different actions to score the strength of a customer interaction with an item. For example, a customer purchasing an item may be a stronger signal than a customer placing an item in a shopping cart, which in turn may be a stronger signal than a customer clicking on an item. The online concierge system generates a ranking of candidate promoted items based on the search query and using the click through rate model. Based on the ranking, the online concierge system displays promoted items along with the organic search results to the customer.