6. Increase Conversion? Optimize Revenues? How?
● Predict consumer’s preferred products and willingness to
pay from online activity and transactions history
● Predict product price elasticity (promotion impact on sales)
● Promote the right product, to the right consumer, at the
right price, in. the right time
1
2
3
7. How?
● Deep learning and context aware recommender system
● Leveraging consumers clickstream in website and purchase history to
detect trends and dynamically predict consumption intent and products
price elasticity
● Smart promotions
8. Personal Promotion
● The maximum price a consumer is willing to pay (WTP) for a
product varies among consumers
● It is possible to implicitly model consumers willingness to pay
from transactions history
● Incorporating consumers willingness to pay in a recommender
system will improve recommendations effectiveness
● Goal: improve recommendations, promotion optimization
9. Experiment: eBay Market Place*
Different Prices by Different Sellers Simultaneously
● 6 Months of transactions and bids
historical data
● Free shipping, US only
● Sellers differ by their reputation
● “Buy now” and auctions
● Per transaction: date, consumer id,
product id, bid value/transaction
price, seller id, seller reputation
10. Personal WTP* Modelling
Same Distribution for All Consumers, Different Parameters
* WTP – Maximum price a consumer is willing to pay
• WTP generic distribution curve per product
• Complementary cumulative curve is the demand curve.
• Personal Transaction price is the personal WTP distribution median.
• Context Aware Recommender System (CARS) method is used to
predict WTP of unseen products, taking into account seller’s reputation
14. Matrix Factorization
* Koren, Y., Bell, R. and Volinsky, C., 2009. Matrix factorization techniques
for recommender systems. Computer, 42(8).
R
V
M
d
x Q d
N
15. Tensor Factorization*
A matrix for
every context variable
Pros: accuracy
Cons: many parameters to learn (small
datasets, computational challenge)
16. Context Aware Matrix Factorization (CAMF)*
* Baltrunas, L., Ludwig, B. and Ricci, F., 2011. Matrix factorization techniques
for context aware recommendation. In Recsys
Pros: computation time
Cons:
- User – context
- No neighborhood
contribution for the interaction
parameters
19. Results: WTP Prediction
Good Accuracy, Incorporating Seller’s Reputation Improves Prediction Accuracy
Average Bid Price = $28, Average Transaction Price = $47
Matrix Factorization (MF), Context Aware Matrix Factorization* (CAMF)
Seller’s reputation was modeled as contextual variable
WAPE: Weighted absolute percentage error
20. Results: Consumption Prediction
Incorporating Seller’s Reputation and Personal Demand Provides Best Results
Offering Ranking = Product Consumption (CARS) * Personal Demand
Implicit feedback. 10 Offerings with highest ranking vs. actual consumed transactions
21. Conclusions
● WTP varies among consumers, it is possible to
implicitly model consumers WTP
● Incorporating personal WTP in a recommender
system improves recommendation accuracy
Promotion Optimization
22. Neural Collaborative Filtering For CARS*
The Consumer Dilemma: Higher Reputation or Lower Price?
* Submitted to UMAP 2019
●
24. Conclusions
• Price sensitivity varies among consumers and products
• It is possible to implicitly model consumers price sensitivity based on
transactions history
• Incorporating personal price sensitivity in a recommender system improves
recommendation accuracy
• Additional features should be incorporated
• CARS is an effective mechanism for this purpose
25. Promotion Planning Optimization
Scenario: Each week ~3% of the products are promoted.
Data: 6 months of clickstream data, purchase history and products catalog.
Train 2x3 months, test on last 2 weeks.
15,000 products in catalog.
Goal: Optimize campaign profits:
promoted quantity * (promotion price – cost) – base quantity * (regular price – cost)
29. Approach
Item
Embedding
RNN
Item Similarity
& Analogy
Clicks
Prediction
* Greenstein-Messica, A., Rokach, L. and Friedman, M., 2017, March. Session-based recommendations using item
embedding. In Proceedings of the 22nd International Conference on Intelligent User Interfaces (pp. 629-633). ACM.
30. Item Embedding for Session Based Recommendations
Data: 2 weeks of clickstream data
and purchase history. Train on 13
days, test on last day.
Goal: recommend relevant
products following first 3 clicks.
Optimize the number of clicked
products which are recommended.
Results: 15% higher
match when
recommending 10
products, 40% clicks
matching