Modeling Seller Listing Strategies

1
Modeling Seller Listing Strategies

• Quang Duong
• University of Michigan
• Neel Sundaresan Nish Parikh Zeqiang
  Shen
• eBay Research Labs

2
Motivation Modeling eBay Sellers Activities

• A majority of eBay sellers are individuals or
  small sale operations (heterogeneous)
• eBay platform provides a wide variety of options
  for listing for-sale item

3
Goal

• Construct a behavior model
• captures seller listing activities
• incorporates historical data and sale
  competitions
• across different product groups/markets
• Domain eBay

4
Applications

• Identify and foster good (listing) practices
  advise and suggest good practices to average
  sellers.
• Assist market design
• For example, eBay platform changes how changes
  impact sellers strategies

5
Related work

• Benefits of Buy it now Anderson et al. 2004
• Clustering sellers Pereira et al. 2009
• Statistical models of agent listing strategies
  Anderson et al. 2007
• Our model incorporates
• dynamic elements
• interactions among sellers

6
Overview
7
Data Processing
Seller ID Product ID Start Date End Date Price Title Shipping
ABC 10 03/20/2009 04/10/20009 100 Silver Nano iPod Apple 0

• Product Clustering
• Need to group listings of the same product
• Use a catalog match each listing to a product in
  the catalog
• Match product name and brand
• Count the number of matched words between
  products catalog description and listings title

8
Data Processing (cont.)

• Data summarization
• Assume sellers adjust their listings in 1-week
  intervals.
• For each 1-week interval, each product and each
  seller
• Average price
• Relative average price
• Number of listings
• (Percentage of free-shipping listings)
• (Percentage of featured listings)
• Product category seller adopt the same strategy
  for products in the same product category
• For example, product black/silver iPhones
  product category iPhone

9
Markov ModelState and Action Representations
State price relative price number of listings
shipping feature
State price (low, med, high), relative price
(low,med,high), number of listings
(low,med,high), shipping (free,not free),
feature (yes,no)
Action Adjust price Adjust number of
listings Adjust shipping cost Adjust feature
selections

• Assumptions
• Markov property only dependent on the immediate
  state (relaxed later)

10
State-Action Model
State price, relative price, number of
listings, shipping, feature Past action
Action Adjust price Adjust number of
listings Adjust shipping cost Adjust feature
selections
Probability Pr(actionstate)
11
Model Learning and Evaluation

• Learning
• Given training data D, learn model Ms
  transition
• Pr(actionstate) Each data point is computed over
  all listings for one product (in one particular
  product category) in a week for a particular
  seller.
• Evaluation
• Given testing data D, compute the log likelihood
  of D with M
• L(M)avg(log(Pr(actionstate))
• Given two models M1 and M2
• L(M1,M2) L(M1) / L(M2) (smaller than 1 means
  M1 is better than M2)
• Final measure 1 - L(M1,M2) ? How much M1 is
  better than M2.

12
Empirical Study

• Examine activities of the best performing seller
  (S0), second best seller (S1), and an average
  seller (S2).
• 3 months worth of data (2/3 for training, 1/3 for
  testing)
• Three product categories charger, battery and
  screen protector (for iPhones)

13
Comparison with the Baseline Semi-uniform Model

• Semi-uniform model (M0)
• Pr(do-nothingstate) is 50
• other actions are randomly uniformly chosen.
• Results for top seller S0 and second-best S1
• Sellers do adopt strategies for their listings

Charger Battery Screen protector
MS0 vs M0 77.9 69.8 77.4
MS1 vs M0 67.1 62.8 57.7
14
Comparison with the History-independent Model

• History-independent model (Mh)
• does not incorporate the last action
• Results for top seller S0
• There are benefits of including information about
  last actions in capturing listing strategies

Charger Battery Screen protector
Ms0 vs Mh 76.1 67.9 61.2
15
Cross-product Analysis

• For seller S0, across different product
  categories
• M1 D1(D2) model trained on product category
  1s data, tested on product category 1(2)s data
• The top seller appears to execute relatively
  different strategies for different product
  categories.

M1 vs M2 D 1 M2 vs M1 D 2
Charger vs Battery 30.1 25.3
Charger vs Screen protector 36.9 22.1
Battery vs Screen protector 32.7 40.6
16
Cross-seller Analysis

• Compare different sellers strategies for the
  same product categories
• The best and second-best sellers have similar
  strategies in the two product categories charger
  and battery, but different strategies for the
  screen protector.
• The top seller and the average seller diverge
  significantly for both charger and screen
  protector

Charger Battery Screen protector
Ms0 vs Ms1 D s0 10 5.6 45
Ms0 vs Ms2 D s0 69 N/A 60
17
Sale-through Rate and Average Revenue Analysis

• We want to compare the effectiveness of seller 0
  and seller 2s strategies
• Sale-through rate
• Average revenue
• Challenge listings created at time t may affect
  sales of previously created listings
• Solution
• Listings sold lt 2 weeks after posted are counted
  as the original actions effect
• Listing sold gt 2 weeks are counted as the newest
  actions effect

18
Conclusions

• Contributions
• Introduce a model that captures sellers listing
  activities, accommodates probabilistic reasoning
  about their behavior, and enables the inclusion
  of historical information
• demonstrate the application of our model in
  comparing listing strategies from different
  sellers across different product categories