E-Commerce

1
E-Commerce
2
Outline

• Introduction
• Customer Data on the Web
• Automated Recommender Systems
• Networks and Recommendations
• Web Path Analysis for Purchase Prediction

3
Introduction

• Some Motivating Questions
• Can we design algorithms to help recommend new
  products to visitors based on their browsing
  behavior?
• Can we better understand factors influencing how
  customers make purchases on a website?
• Can we predict in real time who will make
  purchases based on their observed navigation
  patterns?

4
Customer Data on the Web

• Data collection on client, server sides and
  anywhere in between
• Goal determine who is purchasing what products
• Tracking customer data
• Web logs, E-Commerce logs, cookies, explicit
  login
• Data then used to provide personalized content to
  site users to
• Assist customers in locating their target
  selections
• Encourage customers to make certain selections

5
Automated Recommender Systems

• Problem framed in two ways
• Users vote for pages/items (binary)
• Users rank pages/items (multivalued)
• Results are captured in a generally sparse matrix
  (users x items)
• Complication no votes can occur because users do
  not vote on items they do not like (Breeze, et al
  1998)
• Ignored by most recommender systems

6
Automated Recommender Systems
7
Evaluating Recommender Systems

• Cautions in data interpretation
• Users may purchase items regardless of
  recommendations
• Users may also avoid purchases they might have
  made based on recommendations
• Approaches to recommender algorithms
• Nearest-neighbor
• Model-based collaborative filtering
• Others?

8
Nearest-Neighbor Collaborative Filtering

• Basic principle utilize users vote history to
  predict future votes/recommendations
• Find most similar users to the target user in the
  training matrix and fill in the target users
  missing vote values based on these
  nearest-neighbors
• A typical normalized prediction scheme
• goal predict vote for item j based on
  other users, weighted towards those with
  similar past votes as target user a

9
Nearest-Neighbor Collaborative Filtering

• Another challenge defining weights
• What is the most optimal weight calculation to
  use?
• Requires fine tuning of weighting algorithm for
  the particular data set
• What do we do when the target user has not voted
  enough to provide a reliable set of
  nearest-neighbors?
• One approach use default votes (popular items)
  to populate matrix on items neither the target
  user nor the nearest-neighbor have voted on
• A different approach model-based prediction
  using Dirichlet priors to smooth the votes (see
  chapter 7)
• Other factors include relative vote counts for
  all items between users, thresholding, clustering
  (see Sarwar, 2000)

10
Nearest-Neighbor Collaborative Filtering

• Structure based recommendations
• Recommendations based on similarities between
  items with positive votes (as opposed to votes of
  other users)
• Structure of item dependencies modeled through
  dimensionality reduction via singular value
  decomposition (SVD) aka latent semantic indexing
  (see chapter 4)
• Approximate the set of row-vector votes as a
  linear combination of basis column-vectors
• i.e. find the set of columns to least-squares
  minimize the difference between the row
  estimations and their true values
• Perform nearest-neighbor calculations to project
  predictions for all items

11
Model Based Collaborative Filtering

• Recommendations based on a model of
  relationships between items based on historical
  voting patterns in the training set
• Better performance than nearest-neighbor analysis
• Joint distribution modeling
• Uses one model as basis for predictions
• Conditional distribution modeling
• A model for each item predicting future vote
  based on votes for each of the other items

12
Model Based Collaborative Filtering

• Joint distribution modeling A practical approach
• Model joint distribution as a finite mixture of
  simpler distributions
• Additional simplification is achieved by assuming
  that votes are independent of others within a
  component
• Limitation assumes that users can be described
  with one model of the K mixture components
• Hoffman and Puzicha (1999) propose a workaround
  asserting that each row of votes represents up to
  K mixture components, rather than a single
  component

13
Model Based Collaborative Filtering

• Another limitation all predictions are based on
  the (static) training set
• Conditional distribution modeling
• Better results by creating a model for each item
  conditioned on the others rather than using a
  single joint density model
• Decision trees Heckerman et al. (2000)
• Greedy approach to approximate tree structure
• Predictions are made for each item not purchased
  or visited
• Performance
• Accuracy nearly equal to Bayesian networks
• Offline memory usage significantly less than
  Bayesian networks
• Offline computation time complexity better than
  Bayesian networks

14
Model-Based Combining of Votes and Content

• Combine content-specific information with other
  information (e.g. structure, vote)
• Useful for determining item similarity (Mooney
  and Roy 2000) and creating user models
• Useful when there is no vote history
• Implementation (Popescul et al 2000)
• Extension of (Hoffman and Puzicha 1999)
• Joint density is determined assuming a hidden
  latent variable making users, documents, and
  words conditionally independent i.e.

15
Model-Based Combining of Votes and Content

• The hidden variable represents multiple (hidden)
  topics of a document
• Conditional probabilities of the hidden parameter
  are calculated using EM
• Sparsity still remains a problem for
  content-based modeling

16
Challenges

• Noisy Data
• The same user may use multiple IP
  addresses/logins
• Different users may use the same IP address/login
• Privacy
• No cookies!
• Changing user habits
• Previous history may not accurately predict
  present purchase selection
• Continuous updating of user activities

17
Networks Recommendation

• Word-of-Mouth
• Needs little explicit advertising
• Products are recommended to friends, family,
  co-workers, etc.
• This is the primary form of advertising behind
  the growth of Google

18
Email Product Recommendation

• Hotmail
• Very little direct advertising in the beginning
• Launched in July 1996
• 20,000 subscribers after a month
• 100,000 subscribers after 3 months
• 1,000,000 subscribers after 6 months
• 12,000,000 subscribers after 18 months
• By April 2002 Hotmail had 110 million subscribers

19
Email Product Recommendation

• What was Hotmails primary form of advertising?
• Small link to the sign up page at the bottom of
  every email sent by a subscriber
• Spreading Activation
• Implicit recommendation

20
Spreading Activation

• Network effects
• Even if a small number of people who receive the
  message subscribe (0.1), the service will
  spread rapidly
• This can be contrasted with the current practice
  of SPAM
• SPAM is not sent by friends, family, co-workers
• No implicit recommendation
• SPAM is often viewed as not providing a good
  service

21
Modeling Spreading Activation

• Diffusion Model
• Montgomery (2002)
• Applied models used in marketing literature, Bass
  (1969) to the hotmail phenomena
• Similar word-of-mouth networks used in selling
  consumer electronics such as refrigerators and
  televisions
• We want to predict at time t how many individuals
  k(t) will adopt the product out of a population
  of N possible adopters

22
Modeling Spreading Activation

• Diffusion Model
• Two ways individuals will subscribe
• Direct Advertising
• At time t, N k(t) individuals have not
  subscribed
• a 0 percent of these individuals will subscribe
  due to direct advertising
• Word-of-Mouth
• At time t, there are k(t)(N k(t)) possible
  connections between subscribers and
  non-subscribers
• ß 0 percent of these connections will cause a
  non-subscriber to subscribe

23
Modeling Spreading Activation

• Combine these and we get the following
  expression
• Solve this and we get

24
Modeling Spreading Activation
25
Modeling Spreading Activation
26
Modeling Spreading Activation

• Diffusion Model
• This does not completely model the what actually
  occurred
• However, it is simple and provides a lot of
  interesting (useful) information
• Other work
• Domingos Richardson (2001) Markov Random Field
  Model
• Daley Gani (1999) various deterministic and
  stochastic models

27
Purchase Prediction

• We want to predict whether or not a shopper will
  make a purchase
• We know demographics
• We know page view patterns
• Can we accurately predict if the user will make a
  purchase or not?

28
Purchase Prediction

• Li et al. (2002)
• Study 1160 shoppers at www.barnesandnoble.com
  between April 1 and April 30, 2002
• The data was collected client side so they knew
  exactly what pages were displayed to the user
• They also knew the demographics (predominantly
  well-educated and affluent)

29
Purchase Prediction

• Li et al. (2002)
• There were 14,512 page views which they divided
  into 1659 sessions
• Mean 8.75
• Median 5
• Standard deviation 16.4
• Min 1
• Max 570
• 7 of sessions contained a purchase

30
Purchase Prediction

• Li et al. (2002)
• Divided the pages into 8 classes
• Home (H), main page
• Account (A), account information pages
• List (L), pages with lists of items
• Product (P), page with a single item
• Information (I), informational pages (shipping
  etc.)
• Shopping cart (S)
• Order (O), indicates a completed order
• Entry or Exit (E), entering or leaving the site

31
Purchase Prediction

• Li et al. (2002)
• Each session was represented by a string of the
  form I H H I I L I I E
• A session containing an O is considered having
  made a purchase
• The average length of a session with a purchase
  was 34.5 and without was only 6.8

32
Purchase Prediction

• Markov transition matrix
• For sessions with no purchase

33
Purchase Prediction

• Li et al. (2002)
• They did several models based on this data
• Tested on predicting next page and predicting a
  purchase
• Best models 64 accurate at predicting next page
• After 2 page views the best models predicted 12
  true positives and 5.3 false positives
• After 6 page views 13.1 true positives and 2.9
  false positives