Product Recommendation Model Using Semantic Embeddings for Cold Start

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing product recommendation systems face challenges in recommending related products for items with insufficient user history, leading to biased recommendations and inefficiencies in handling large-scale data, particularly in the shopping domain.

Innovation Solution

A related product recommendation model is developed using a natural language inference task to classify relations between sentences, leveraging a contrastive pair dataset and a Siamese network to generate embeddings that capture dynamic correlations between products, incorporating both history and content information.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If history-based recommendation algorithms are used, then recommendation accuracy for items with sufficient history is improved, but recommendation capability for cold items deteriorates

Engineering Contradiction:
Improverecommendation accuracyVSAvoidrecommendation capability for cold items
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent introduces a natural language processing intermediary that processes product titles and descriptions to generate semantic embeddings. This intermediary enables cold items without historical data to be represented and recommended based on their content semantics, bridging the gap between history-based accuracy and cold item capability

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent transforms product information from traditional structured parameters (ID, category) to semantic parameters (embeddings from NLP models). This parameter transformation allows the system to leverage semantic similarity rather than relying solely on historical interaction data, enabling effective cold item recommendation

Inventive Principle:
Principle #35Parameter changes

2Speed

If traditional recommendation models are used, then processing speed for simple queries is improved, but handling capability for large-scale data deteriorates

Engineering Contradiction:
Improveprocessing speedVSAvoidhandling capability for large-scale data
Core Design Contradiction:
SpeedVSQuantity of substance

Solution Approach 1:

The patent performs preliminary processing by pre-computing semantic embeddings for all products using NLP models and storing them in advance. This preliminary action transforms raw text data into structured semantic representations, enabling efficient similarity search and recommendation generation without heavy real-time processing for large-scale data queries

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent creates semantic copies (embeddings) of product information that capture the essential meaning without requiring the full original text data during recommendation generation. These copied representations enable fast comparison and matching across large product catalogs while reducing computational overhead

Inventive Principle:
Principle #26Copying

3Adaptability or versatility

If history-based recommendation systems are used, then personalization for users with extensive history is improved, but bias toward popular items deteriorates

Engineering Contradiction:
Improvepersonalization capabilityVSAvoidrecommendation bias
Core Design Contradiction:
Adaptability or versatilityVSObject-generated harmful factors

Solution Approach 1:

The patent segments the recommendation approach into two independent components: user-specific historical behavior analysis and item-level semantic content analysis. By separating these functions, the system can personalize recommendations based on user history while simultaneously considering the intrinsic semantic value of items, thereby reducing bias toward popular items and enabling discovery of relevant but less popular products

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4270291A1Method and computer device to recommend related product technical field
Publication Date: 2023.11.01 NAVER CORP
  • EP4270291A1 patent drawingFigure 1
  • EP4270291A1 patent drawingFigure 2
  • EP4270291A1 patent drawingFigure 3

AI summary

Disclosed are a method, a computer device, and a non-transitory computer-readable recording medium to recommend a related product. The method to recommend the related product may include constructing a contrastive pair dataset for a natural language inference task using content information related to a product based on history of the product in a shopping domain; and generating a product recommendation model to apply to the shopping domain through pre-training of a language model for the natural language inference task using the contrastive pair dataset.