Dynamic Merchant Similarity Grouping via Vectorization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current systems for identifying similar merchants in delivery platforms rely on explicit coding and purchase history, failing to leverage implicit user preferences and capture latent user preferences effectively, especially in real-time on-demand deliveries of perishable goods.
Innovation Solution
A neural network-based system that converts merchant and user data into vector representations, using skip-gram modeling to generate dynamic similarity predictions by analyzing historical browsing sessions, allowing for implicit preference capture and targeted recommendations without explicit encoding.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If explicit coding and purchase history are used to identify similar merchants, then the system can provide basic merchant recommendations, but it fails to capture implicit user preferences and latent preferences effectively
Solution Approach 1:
The patent replaces traditional explicit coding and purchase history analysis with a neural network-based vectorization system. The neural network automatically processes browsing session data to generate merchant and user vectors, substituting manual feature engineering and explicit data collection with an automated learning system that can infer implicit preferences from behavioral patterns.
Solution Approach 2:
The patent introduces vector representations as an intermediary between raw browsing data and recommendation outputs. These vectors serve as a mediator that transforms complex, unstructured browsing session data into a form that can be efficiently processed by the neural network to identify latent preferences and generate accurate recommendations.
2Adaptability or versatility
If traditional merchant similarity identification methods are used, then the system structure remains simple, but the system cannot provide personalized recommendations based on implicit user preferences
Solution Approach 1:
The patent creates a universal vectorization framework that can handle multiple types of data (merchant attributes, user browsing behavior, item characteristics) through a single neural network architecture. This multi-functional system processes diverse inputs and generates recommendations across different contexts, making the complex system adaptable to various recommendation scenarios.
Solution Approach 2:
The patent transforms the recommendation problem from categorical matching to continuous vector space operations. By converting discrete merchant and user data into continuous vector representations, the system can apply mathematical operations (cosine similarity, dot products) to measure relationships, fundamentally changing the parameters from categorical to continuous and enabling more nuanced personalization.
3Measurement precision
If browsing session data is vectorized and processed through neural networks, then implicit user preferences can be captured effectively, but the computational complexity and data processing requirements increase
Solution Approach 1:
The patent performs preliminary vectorization of merchant and user data during offline training phases, pre-computing vector representations that can be quickly retrieved and compared during online recommendation generation. This preliminary action separates the computationally intensive training process from the real-time recommendation process, reducing online computational requirements.
Solution Approach 2:
The patent creates simplified vector copies of complex merchant and user profiles that capture essential preference patterns. These vector representations are lightweight copies that can be rapidly processed and compared without requiring access to the full original data, significantly reducing computational overhead during recommendation generation.
Data Source
AI summary
Provided are various mechanisms and processes for generating dynamic merchant similarity predictions. In one aspect, a system is configured for receiving historical datasets that include a series of merchants from historical browsing sessions generated by one or more users. The merchants are converted into corresponding vector representations for training a predictive model to output associated merchants based on a generated weighted vector space. Once sufficiently trained, data from a new browsing session may be received, which may include a target merchant. The target merchant is input into the predictive model as a vector to output one or more context merchants having vectors with the highest cosine similarity value to the target merchant vector. Selected context merchants may then be transmitted to the user device as targeted merchant suggestions in the new browsing session. The predictive models may be continuously trained using data received from subsequent browsing sessions.