ݺߣ

ݺߣShare a Scribd company logo

poster

•
•
Q
Qi Zhao

This document proposes adding personalized promotion to e-commerce recommendation systems. It introduces a novel willingness-to-pay elicitation procedure applicable to e-commerce. The method models consumer willingness-to-pay for products as a regression function of interaction features between consumers and products. An experiment with 120k skin care products from Amazon and 483 records from 79 subjects was used to evaluate the approach. Results showed the potential for personalized pricing and promotion to be useful in e-commerce recommendation.

1 of 1
Download to read offline
E-commerce Recommendation with Personalized Promotion Qi Zhao1 , Yi Zhang2 , Daniel Friedman3 1 Technology and Information Management 2 Computer Science 3 Econonics Experiment Setting Motivation Existing e-commerce recommendation algorithms treat the properties, including the price, of the product fixed. Price plays an important role in consumer purchase decision and can be personalized according to individual preference. The key is to estimate consumer's Willingness-to-Pay (WTP) The major contributions of the proposed work are, Our work is the first attempt to add personalized promotion into e-commerce recommender systems. We introduced a novel WTP elicitation procedure that is applicable to e-commerce system. We have various interesting observations that could be very useful for further research on personalized pricing or promotion in e-commerce. Method Overview WTP Modeling Let denote consumer 's WTP on product , and denote the interaction feature between and , is modeld as regression function w.r.t , The model parameter vector can be inferred by minimizing the emperical loss on the training data as follows, , where L is loss function which measures the goodness of fit. In our experiment, gradient boosting trees (GBT) and linear regression were adopted as regression function and quadratic function is used for L. 120k skin care products from Amazon. We recuirt experiment subjects from Amazon Mechnical Turk. We pay each subject $0.5 for finishing the task. Subject bids at least 5 products. We set lottery $100. We collected 483 records from 79 subjects. Evaluation Metric root Mean Squared Error (RMSE) for regression error. Conversion rate for the occurance of purchase. Seller gross revenue for effectiveness of WTP prediction Result first ranked brand most selected brand list price number of reviews brand variety discount Amazon rec. rank average rating user rank switch brand 0 5 10 15 20 25 30 0 50 100 150 0.550.600.650.70 f(pr) WTP w.r.t Amazon list price 0.0 0.2 0.4 0.6 0.8 1.0 0.580.600.620.64 f(ar_f) WTP w.r.t average rating 0.0 0.2 0.4 0.6 0.8 0.600.650.70 f(prmt_f) WTP w.r.t Amazon promotion 0 1 0.6180.6200.6220.6240.626 f(switch_brand_b) WTP w.r.t Amazon list price

More Related Content

poster

  • 1. E-commerce Recommendation with Personalized Promotion Qi Zhao1 , Yi Zhang2 , Daniel Friedman3 1 Technology and Information Management 2 Computer Science 3 Econonics Experiment Setting Motivation Existing e-commerce recommendation algorithms treat the properties, including the price, of the product fixed. Price plays an important role in consumer purchase decision and can be personalized according to individual preference. The key is to estimate consumer's Willingness-to-Pay (WTP) The major contributions of the proposed work are, Our work is the first attempt to add personalized promotion into e-commerce recommender systems. We introduced a novel WTP elicitation procedure that is applicable to e-commerce system. We have various interesting observations that could be very useful for further research on personalized pricing or promotion in e-commerce. Method Overview WTP Modeling Let denote consumer 's WTP on product , and denote the interaction feature between and , is modeld as regression function w.r.t , The model parameter vector can be inferred by minimizing the emperical loss on the training data as follows, , where L is loss function which measures the goodness of fit. In our experiment, gradient boosting trees (GBT) and linear regression were adopted as regression function and quadratic function is used for L. 120k skin care products from Amazon. We recuirt experiment subjects from Amazon Mechnical Turk. We pay each subject $0.5 for finishing the task. Subject bids at least 5 products. We set lottery $100. We collected 483 records from 79 subjects. Evaluation Metric root Mean Squared Error (RMSE) for regression error. Conversion rate for the occurance of purchase. Seller gross revenue for effectiveness of WTP prediction Result first ranked brand most selected brand list price number of reviews brand variety discount Amazon rec. rank average rating user rank switch brand 0 5 10 15 20 25 30 0 50 100 150 0.550.600.650.70 f(pr) WTP w.r.t Amazon list price 0.0 0.2 0.4 0.6 0.8 1.0 0.580.600.620.64 f(ar_f) WTP w.r.t average rating 0.0 0.2 0.4 0.6 0.8 0.600.650.70 f(prmt_f) WTP w.r.t Amazon promotion 0 1 0.6180.6200.6220.6240.626 f(switch_brand_b) WTP w.r.t Amazon list price