Authors / Chia-Hui Chang and Jun-Hong Lin Presenter / Meng-Lun Wu Decision Support and Profit Prediction for Online Auction Sellers

Outline INTRODUCTION DATA TO SELL OR NOT TO SELL? END-PRICE PREDICTION SOLD PROBABILITY CALIBRATION EXPRIMENTS CONCLUSIONS

INTRODUCTION

DATA

TO SELL OR NOT TO SELL?

END-PRICE PREDICTION

SOLD PROBABILITY CALIBRATION

EXPRIMENTS

CONCLUSIONS

INTRODUCTION Online auction has become a very popular e-commerce transaction type. The main issue is to find an auction setting that could maximize sellers profit is a challenging problem. profits prediction. Ghani and Simmons apply three models for end-price prediction. regression, multi-class classification and multiple binary classification.

Online auction has become a very popular e-commerce transaction type.

The main issue is to find an auction setting that could maximize sellers profit is a challenging problem.

profits prediction.

Ghani and Simmons apply three models for end-price prediction.

regression, multi-class classification and multiple binary classification.

INTRODUCTION (cont.) Some researchers suggest the use of predicted end-price for selling strategy, but they didn’t address how to apply predicted end-price in decision making. This paper proposed the use of cost-sensitive decision making to resolve whether to list a commodity or not. Authors use profit gain over average profit of similar items sold by similar sellers.

Some researchers suggest the use of predicted end-price for selling strategy, but they didn’t address how to apply predicted end-price in decision making.

This paper proposed the use of cost-sensitive decision making to resolve whether to list a commodity or not.

Authors use profit gain over average profit of similar items sold by similar sellers.

DATA Authors collect the digital cameras auction data from eBays, and choose the period of two months in March-April 2006. These data are classified into 3 classes: Seller Features, Items Features, Auction Features

Authors collect the digital cameras auction data from eBays, and choose the period of two months in March-April 2006.

These data are classified into 3 classes:

Seller Features, Items Features, Auction Features

DATA (cont.)

DATA (cont.)

Two tasks

TO SELL OR NOT TO SELL? This paper have three approaches of decision criteria: Sold probability Based Approach When the predicted sold probability is greater than 50%. P(c=1|x) > P(c=-1|x) End-price Based Approach When the predicted end-price is higher than the average end-price of similar items y(x) > Avgy(x)

This paper have three approaches of decision criteria:

Sold probability Based Approach

When the predicted sold probability is greater than 50%.

P(c=1|x) > P(c=-1|x)

End-price Based Approach

When the predicted end-price is higher than the average end-price of similar items

y(x) > Avgy(x)

TO SELL OR NOT TO SELL? (cont.) Expected profit Based Approach When the expected profit gain for suggesting sell is greater than zero. P(c=1|x)[Profit(x) – AvgP(x)]>P(c=-1|x)[lc(x)+uc]

Expected profit Based Approach

When the expected profit gain for suggesting sell is greater than zero.

P(c=1|x)[Profit(x) – AvgP(x)]>P(c=-1|x)[lc(x)+uc]

END –PRICE PREDICTION There are two types of auction end-price predictions Static : Multiple Binary Classification Dynamic : Sample Selection Bias Correction Ghani and Simmons proposed Multiple Binary Classification prediction method with neural network algorithm get the better accuracy. Authors use multiple binary classifiers with two-class estimation to predict the end-price for auction listing items.

There are two types of auction end-price predictions

Static : Multiple Binary Classification

Dynamic : Sample Selection Bias Correction

Ghani and Simmons proposed Multiple Binary Classification prediction method with neural network algorithm get the better accuracy.

Authors use multiple binary classifiers with two-class estimation to predict the end-price for auction listing items.

TWO–CLASS ESTIMATION 47.5

Sample Selection Bias Correction Sample selection bias problem has been studied in econometrics and statistics. There are four kinds of sample selection bias: Complete Independent Feature Bias Class Bias Complete Bias

Sample selection bias problem has been studied in econometrics and statistics.

There are four kinds of sample selection bias:

Complete Independent

Feature Bias

Class Bias

Complete Bias

SOLD PROBABILITY CALIBATION To estimation the sold probability for an item, we can use any binary classification to train a model for predicting whether an item would be sold (c=1) or not (c=-1). In this paper, we use SVM as our classifier. SVM produces an uncalibrated value that is not a probability. To yield well calibrated probability, we adopt Platt Scaling to transfer a value f(x i ) into probability P(c=1|x i ).

To estimation the sold probability for an item, we can use any binary classification to train a model for predicting whether an item would be sold (c=1) or not (c=-1).

In this paper, we use SVM as our classifier.

SVM produces an uncalibrated value that is not a probability.

To yield well calibrated probability, we adopt Platt Scaling to transfer a value f(x i ) into probability P(c=1|x i ).

EXPERIMENTS 70% data as training data and 30% as testing data. The experiments divided into three parts: The performance on item sold prediction The accuracy of end-price prediction Compares three performance of three approaches Probability-based End-price based Expected profit based

70% data as training data and 30% as testing data.

The experiments divided into three parts:

The performance on item sold prediction

The accuracy of end-price prediction

Compares three performance of three approaches

Probability-based

End-price based

Expected profit based

Accuracy of Sold Probability Authors use SVM as our classifier for predicting. The accuracy of prediction whether an item will be sold I about 93.5% and 76.2% for auction listing and BuyItNow listing.

Authors use SVM as our classifier for predicting.

The accuracy of prediction whether an item will be sold I about 93.5% and 76.2% for auction listing and BuyItNow listing.

END-price Prediction To use multiple binary classifiers for end-price prediction, we need to decide the reference value for each binary classifier. Authors exclude the highest and lowest end-prices, and use 10% window of the average end-price as the interval size.

To use multiple binary classifiers for end-price prediction, we need to decide the reference value for each binary classifier.

Authors exclude the highest and lowest end-prices, and use 10% window of the average end-price as the interval size.

Profit Gain Comparison In order to validate that end-price alone does not guarantee high profit due to low sold probability, authors use profit gain over other sellers to compare the three approaches. Sell all approach as the baseline for comparison.

In order to validate that end-price alone does not guarantee high profit due to low sold probability, authors use profit gain over other sellers to compare the three approaches.

Sell all approach as the baseline for comparison.

Auction Listing

BuyItNow Listing

CONCLUSIONS A critical question for online auction sellers is to find an auction setting. Authors apply machine learning algorithms for end-price prediction and sold probability estimation. For end-price prediction, since we can only use old items to train the classifiers, the model could under-estimate the end-price for unsold items.

A critical question for online auction sellers is to find an auction setting.

Authors apply machine learning algorithms for end-price prediction and sold probability estimation.

For end-price prediction, since we can only use old items to train the classifiers, the model could under-estimate the end-price for unsold items.

CONCLUSIONS (cont.) As shown in the experiments, approaches that don’t involve the consideration of sold probability have deteriorating profit as ultra cost increase. This proves our argument that sold probability should also be considered for profit maximization. The approach based on both probability and end-price have the best performances among all other approaches.

As shown in the experiments, approaches that don’t involve the consideration of sold probability have deteriorating profit as ultra cost increase.

This proves our argument that sold probability should also be considered for profit maximization.

The approach based on both probability and end-price have the best performances among all other approaches.