Utilizing Marginal Net Utility for Recommendation in E-commerce
Utilizing Marginal Net Utility for Recommendation in E-commerce Jian Wang & Yi Zhang Informa(on Retrieval and Knowledge Management Lab University of California, Santa Cruz
1
Recommenda(on based on Predicted Ra(ng is Not Good for Consumers • Will you like it? • Tradi(onal recommender systems recommend item with the highest predicted ra(ng
• Will you purchase it? • Diminishing of return • A ra(onal user will purchase the product with the highest marginal net u(lity
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
2
Total Utility, Marginal Utility Total U(lity
Purchase Quan(ty
Marginal U(lity Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
3
Total Utility, Marginal Utility Total U(lity
Purchase Quan(ty
Marginal U(lity Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
4
Total Utility, Marginal Utility Total U(lity
Purchase Quan(ty
Marginal U(lity Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
5
Total Utility, Marginal Utility Total U(lity
Purchase Quan(ty
Marginal U(lity Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
6
Marginal Net Utility
Marginal Net Utility = Marginal Utility - Price
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
7
Making Consumers Happier
1. Model the consumer behavior based on marginal net u6lity
2. Make recommenda6ons to maximize the marginal net u6lity for each user
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
8
Outline • Mo(va(on • Algorithm – Basic economics: two exis(ng u(lity func(ons studied by economists – Problem seOng – Propose a new u(lity func(on tailored for recommender systems
• Experimental Results • Conclusion Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
9
Basic Economics: Linear Utility Function
• does not capture diminishing of return characteristic • most existing algorithms are implicitly based on this utility function • • • • •
Total utility purchase history product ’s basic utility purchase quantity of product in marginal utility for the addition purchase of
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
10
Basic Economics: Cobb-Douglas Utility
• diminishing of return rate is fixed • basic utility is not personalized • • • • •
Total utility purchase history product ’s basic utility purchase quantity of product in marginal utility for the addition purchase of
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
11
Making Consumers Happier: Recommendation based on Marginal Net Utility Design
Learn
• Design the u6lity func6onal form • Learn the u6lity func6on parameters from user history
• Predict the marginal net u6lity of a product for a user using Predict the func6on learned and user history Rank
• Rank products based on predicted marginal net u6lity
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
12
Problem SeOng
day 1
day 3
day 7
day 10
Time t Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
13
Design the U(lity Func(onal Form CobbDouglas
New
• • • • •
Total utility purchase history product ’s basic utility purchase quantity of product in marginal utility for the addition purchase of
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
14
Design the U(lity Func(onal Form
• only same product will influence the marginal utility New framework to revamp exis6ng recommenda6on algorithms
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
15
Revamp Exis(ng Algorithms Marginal net utility
Take Singular Value Decomposition (SVD) as an example
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
16
Learn the U(lity Func(on Parameters Maximum A Posteriori Estimation
Joint Likelihood
Purchase Likelihood Conditional on the marginal net utility
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
17
Outline
• Mo(va(on • Algorithm • Experimental Results • Conclusion
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
18
Experiment Dataset • More than 5-years purchase history from – 2004-01-01 to 2009-03-08 • 10,399 users, 65,551 products, and 102,915 unique (user, product) pairs • 80% training, 10% validation, and 10% testing
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
19
Experiment Design • Methods to compare - - - same product - similar product • Evaluation Metric
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
20
General Analysis conversion rate@K 0.16 Conversion Rate
0.14 0.12 0.1 0.08
87.39%
0.06 0.04 0.02 0 1
2
3
4
5
K
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
21
Further Analysis • Consider all products – 13.79% repurchase – 90.64% new purchase
• Evaluate orders with the specific purchase type (repurchase or new purchase)
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
22
Further Analysis • The new u(lity func(on achieves significantly beber performance in both tasks • Examples of Repurchases
– diaper, pet food, etc. – Tend to be consumable products
• Examples of New purchases
– computer, cell phone, bed frame, etc – Tend to be durable product – law of diminishing marginal u(lity Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
23
Outline • Mo(va(on • Algorithm • Experimental Results • Conclusion
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
24
Conclusion • Introduce a new framework for recommender system in e-commerce sites • Recommend products with the highest marginal net utility • Take SVD as an example to revamp • Achieve significant improvement in the conversion rate
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
25
Ques(ons?
Utilizing Marginal Net Utility for Recommendation in E-commerce Jian Wang & Yi Zhang Informa(on Retrieval and Knowledge Management Lab University of California, Santa Cruz
26
1
Recommenda(on based on Predicted Ra(ng is Not Good for Consumers • Will you like it? • Tradi(onal recommender systems recommend item with the highest predicted ra(ng
• Will you purchase it? • Diminishing of return • A ra(onal user will purchase the product with the highest marginal net u(lity
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
2
Total Utility, Marginal Utility Total U(lity
Purchase Quan(ty
Marginal U(lity Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
3
Total Utility, Marginal Utility Total U(lity
Purchase Quan(ty
Marginal U(lity Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
4
Total Utility, Marginal Utility Total U(lity
Purchase Quan(ty
Marginal U(lity Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
5
Total Utility, Marginal Utility Total U(lity
Purchase Quan(ty
Marginal U(lity Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
6
Marginal Net Utility
Marginal Net Utility = Marginal Utility - Price
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
7
Making Consumers Happier
1. Model the consumer behavior based on marginal net u6lity
2. Make recommenda6ons to maximize the marginal net u6lity for each user
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
8
Outline • Mo(va(on • Algorithm – Basic economics: two exis(ng u(lity func(ons studied by economists – Problem seOng – Propose a new u(lity func(on tailored for recommender systems
• Experimental Results • Conclusion Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
9
Basic Economics: Linear Utility Function
• does not capture diminishing of return characteristic • most existing algorithms are implicitly based on this utility function • • • • •
Total utility purchase history product ’s basic utility purchase quantity of product in marginal utility for the addition purchase of
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
10
Basic Economics: Cobb-Douglas Utility
• diminishing of return rate is fixed • basic utility is not personalized • • • • •
Total utility purchase history product ’s basic utility purchase quantity of product in marginal utility for the addition purchase of
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
11
Making Consumers Happier: Recommendation based on Marginal Net Utility Design
Learn
• Design the u6lity func6onal form • Learn the u6lity func6on parameters from user history
• Predict the marginal net u6lity of a product for a user using Predict the func6on learned and user history Rank
• Rank products based on predicted marginal net u6lity
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
12
Problem SeOng
day 1
day 3
day 7
day 10
Time t Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
13
Design the U(lity Func(onal Form CobbDouglas
New
• • • • •
Total utility purchase history product ’s basic utility purchase quantity of product in marginal utility for the addition purchase of
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
14
Design the U(lity Func(onal Form
• only same product will influence the marginal utility New framework to revamp exis6ng recommenda6on algorithms
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
15
Revamp Exis(ng Algorithms Marginal net utility
Take Singular Value Decomposition (SVD) as an example
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
16
Learn the U(lity Func(on Parameters Maximum A Posteriori Estimation
Joint Likelihood
Purchase Likelihood Conditional on the marginal net utility
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
17
Outline
• Mo(va(on • Algorithm • Experimental Results • Conclusion
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
18
Experiment Dataset • More than 5-years purchase history from – 2004-01-01 to 2009-03-08 • 10,399 users, 65,551 products, and 102,915 unique (user, product) pairs • 80% training, 10% validation, and 10% testing
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
19
Experiment Design • Methods to compare - - - same product - similar product • Evaluation Metric
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
20
General Analysis conversion rate@K 0.16 Conversion Rate
0.14 0.12 0.1 0.08
87.39%
0.06 0.04 0.02 0 1
2
3
4
5
K
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
21
Further Analysis • Consider all products – 13.79% repurchase – 90.64% new purchase
• Evaluate orders with the specific purchase type (repurchase or new purchase)
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
22
Further Analysis • The new u(lity func(on achieves significantly beber performance in both tasks • Examples of Repurchases
– diaper, pet food, etc. – Tend to be consumable products
• Examples of New purchases
– computer, cell phone, bed frame, etc – Tend to be durable product – law of diminishing marginal u(lity Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
23
Outline • Mo(va(on • Algorithm • Experimental Results • Conclusion
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
24
Conclusion • Introduce a new framework for recommender system in e-commerce sites • Recommend products with the highest marginal net utility • Take SVD as an example to revamp • Achieve significant improvement in the conversion rate
Jian Wang & Yi Zhang Utilizing Marginal Net Utility for Recommendation in E-commerce
25
Ques(ons?
Utilizing Marginal Net Utility for Recommendation in E-commerce Jian Wang & Yi Zhang Informa(on Retrieval and Knowledge Management Lab University of California, Santa Cruz
26
