Scholarlink

<p>The market price of a used vehicle is the aggregate result of complex multi-criteria decision-making (MCDM) by prospective buyers. This paper posits that machine learning models can achieve superior predictive accuracy by directly modeling this underlying cognitive process. We contrast two methodologies on a dataset of over 100,000 Moroccan used cars. The classical approach uses state-of-the-art ensemble models on a feature set engineered with traditional encoding. Our novel approach, grounded in Information Integration Theory, replaces four key decision criteria—brand, model, year, and equipment—with a single, LLM-generated “Vehicle Desirability Score.” This score is explicitly designed to emulate an expert’s holistic evaluation, representing a proxy for a consumer’s cognitive utility assessment. Our findings show that models trained with this single, semantically aware feature significantly outperform their counterparts. For the top-performing XGBoost model, using the single “Vehicle Desirability Score” reduced the MAE by 18.4% and the RMSE by 17.6%, while increasing the <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(R^2\)</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mi>R</mi> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation> score from 0.901 to an impressive 0.945. This study contributes to the marketing and analytics literature by demonstrating that using LLMs to model the latent human decision-making process is a powerful new paradigm for feature engineering.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Semantic Feature Engineering: Predicting Used Car Prices by Modeling Multi-criteria Human Decisions with LLMs

  • 期刊论文

  • 发表日期：

    2026年1月7日

查看全文
  • Abdelilah Mhamedi,
  • Mohammed Mghari,
  • Abdelaaziz El Hibaoui

摘要

The market price of a used vehicle is the aggregate result of complex multi-criteria decision-making (MCDM) by prospective buyers. This paper posits that machine learning models can achieve superior predictive accuracy by directly modeling this underlying cognitive process. We contrast two methodologies on a dataset of over 100,000 Moroccan used cars. The classical approach uses state-of-the-art ensemble models on a feature set engineered with traditional encoding. Our novel approach, grounded in Information Integration Theory, replaces four key decision criteria—brand, model, year, and equipment—with a single, LLM-generated “Vehicle Desirability Score.” This score is explicitly designed to emulate an expert’s holistic evaluation, representing a proxy for a consumer’s cognitive utility assessment. Our findings show that models trained with this single, semantically aware feature significantly outperform their counterparts. For the top-performing XGBoost model, using the single “Vehicle Desirability Score” reduced the MAE by 18.4% and the RMSE by 17.6%, while increasing the \(R^2\) R 2 score from 0.901 to an impressive 0.945. This study contributes to the marketing and analytics literature by demonstrating that using LLMs to model the latent human decision-making process is a powerful new paradigm for feature engineering.