商品推荐系统中的效用估计

Utility Estimation in Recommendation System

在零售商的角度,通常希望推荐系统的推荐产品能使商家的收益最大化。在以期望收益最大化为目标的产品组合优化模型中,商品效用是必不可少的参数。论文主要探究推荐系统中商品效用的估计方法,通过评估由商品效用计算得到的商品被点击概率,来验证效用估计的准确性。通过数值试验,将单值排序模型预估的点击概率与通过MNL模型估计商品效用计算的点击概率进行对比,结果证明MNL模型估计的商品效用具备与单值排序模型相当的准确率。此外,论文进一步构建了神经网络模型估计商品效用,并且另外构建了一个直接计算商品点击概率的attention选择模型作为对比。结果证明使用神经网络模型代替MNL来估计商品效用,能够更进一步提升效用预估的准确性。

From the retailer's perspective,it is usually hoped that the recommended products of the recommendation system can maximize the profit of the assortment.In the assortment optimization model aiming at maximizing expected revenues,product utility is an indispensable parameter.This article mainly explores the estimation method of product utility in the recommendation system,and verifies the accuracy of utility estimation by evaluating the click probability of the product calculated by the product utility.Through numerical experiments,the estimated click probability of the point-wise model is compared with the click probability calculated by the product utility estimated by the MNL model.The result proves that the product utility estimated by the MNL model has a good accuracy.In addition,this paper further builds a neural network model to estimate the utility of goods,and additionally builds an attention choice model that directly calculates the click probability of goods as a comparison.The results prove that using neural network model instead of MNL to estimate the utility of the commodity can further improve the accuracy of the utility prediction.