## Cross-sell: A Fast Promotion-Tunable Customer-item Recommendation Method Based on Conditionally Independent Probabilities (2000)

Venue: | In Proceedings of ACM SIGKDD International Conference |

Citations: | 16 - 0 self |

### BibTeX

@INPROCEEDINGS{Kitts00cross-sell:a,

author = {Brendan Kitts and David Freed and Martin Vrieze},

title = {Cross-sell: A Fast Promotion-Tunable Customer-item Recommendation Method Based on Conditionally Independent Probabilities},

booktitle = {In Proceedings of ACM SIGKDD International Conference},

year = {2000},

pages = {437--446},

publisher = {ACM Press}

}

### OpenURL

### Abstract

We develop a method for recommending products to customers with applications to both on-line and surface mail promotional offers. Our method differs from previous work in collaborative filtering [8] and imputation [18], in that we assume probabilities are conditionally independent. This assumption, which is also made in Nave Bayes [5], enables us to pre-compute probabilities and store them in main memory, enabling very fast performance on millions of customers. The algorithm supports a variety of tunable parameters so that the method can address different promotional objectives. We tested the algorithm at an on-line hardware retailer, with 17,400 customers divided randomly into control and experimental groups. In the experimental group, clickthrough increased by +40% (p<0.01), revenue by +38% (p<0.07), and units sold by +61% (p<0.01). By changing the algorithm's parameter settings we found that these results could be improved even further. This work demonstrates the considerable potent...

### Citations

2675 | Fast algorithms for mining association rules
- Agrawal, Srikant
- 1994
(Show Context)
Citation Context ...egory. This particular method of filling in missing values is known in the statistics literature as conditional mean imputation [18]. Formally, let a customer profile x consist of a binary vector x = =-=[0,1]-=-sN where a x si =1 means that customer s purchased/clicked product/web-page i, and a 0 means that the customer did not, and N is the number of variables in the profile 1 . We are trying to predict the... |

1123 |
DB: Statistical Analysis with Missing Data
- RJA, Rubin
- 1987
(Show Context)
Citation Context ...il promotions. Keywords Imputation, cross-sell, collaborative filtering, recommendation 1. INTRODUCTION Most customer recommendation algorithms can be understood as performing some kind of imputation =-=[13]. Some of -=-the customer's interests are known because they have entered "star ratings" or have bought a product, but most are not. The problem of deciding what product to recommend next involves findin... |

1020 | Empirical analysis of predictive algorithms for collaborative filtering. Paper presented at the
- JS, Heckerman, et al.
- 1998
(Show Context)
Citation Context .... Now consider that incremental profit is the conditional probability minus the baseline probability. These two measures are similar in that both are discounting the baseline probability in some way. =-=[2] also foun-=-d that discounting base rating frequencies increased accuracy in predicting interest in test data. Their "inverse user weighting" scheme increased accuracy in all 24 experiments they ran on ... |

524 |
An Algorithmic Framework for Performing Collaborative Filtering
- Herlocker, Konstan, et al.
- 1999
(Show Context)
Citation Context ...en that MV x sj = ( ) i x x x d D si di si " =s= = 1 : } { + = D d dj sj x D x # 1 Other typical imputation algorithms include regression imputation [17,19], the EM algorithm, and hot-deck imputa=-=tion [7,14,4]-=-. Regression imputation selects donor cases in exactly the same way, and then calculates a least squares estimate: [ ] w x x si sjs= + 1 where [ ] dj D di x x ws= -1 # 1 1 This is not the only choice ... |

377 |
JL: Analysis of incomplete multivariate data. Monographs on Statistics and applied Probability 72. Boca Raton USA
- Schafer
- 1997
(Show Context)
Citation Context ...ethod for recommending products to customers with applications to both on-line and surface mail promotional offers. Our method differs from previous work in collaborative filtering [8] and imputation =-=[18]-=-, in that we assume probabilities are conditionally independent. This assumption, which is also made in Nave Bayes [5], enables us to pre-compute probabilities and store them in main memory, enabling ... |

150 |
Multiple Imputation After 18+ Years
- Rubin
- 1996
(Show Context)
Citation Context ...ue". Conditional mean imputation is defined as: Given that MV x sj = ( ) i x x x d D si di si " =s= = 1 : } { + = D d dj sj x D x # 1 Other typical imputation algorithms include regression i=-=mputation [17,19]-=-, the EM algorithm, and hot-deck imputation [7,14,4]. Regression imputation selects donor cases in exactly the same way, and then calculates a least squares estimate: [ ] w x x si sjs= + 1 where [ ] d... |

58 | A non-invasive learning approach to building web user profiles
- Chan
- 1999
(Show Context)
Citation Context ... number can be inverted and interpreted as the number of times lower than random that two items occur. Interestingly, lift is related to the Mutual Information Criterion (MIC) from information theory =-=[3]-=-. MIC is equal to log of lift. We favor the untransformed lift score because it is easier to interpret for the user. 4.1.3 Expected Profit If we assume mutual independence between products, then the e... |

47 | Multiple imputation for multivariate missing-data problems: A data analystâ€™s perspective
- Schafer, Olsen
- 1998
(Show Context)
Citation Context ...ue". Conditional mean imputation is defined as: Given that MV x sj = ( ) i x x x d D si di si " =s= = 1 : } { + = D d dj sj x D x # 1 Other typical imputation algorithms include regression i=-=mputation [17,19]-=-, the EM algorithm, and hot-deck imputation [7,14,4]. Regression imputation selects donor cases in exactly the same way, and then calculates a least squares estimate: [ ] w x x si sjs= + 1 where [ ] d... |

18 | An overview of hot-deck procedures - Ford - 1983 |

8 | Multiple imputation: a primer. Statistical methods in medical research - Schafer - 1999 |

2 |
Estimating the Variance impact of Missing CPS Income Data
- Oh, Scheuren
- 1980
(Show Context)
Citation Context ...en that MV x sj = ( ) i x x x d D si di si " =s= = 1 : } { + = D d dj sj x D x # 1 Other typical imputation algorithms include regression imputation [17,19], the EM algorithm, and hot-deck imputa=-=tion [7,14,4]-=-. Regression imputation selects donor cases in exactly the same way, and then calculates a least squares estimate: [ ] w x x si sjs= + 1 where [ ] dj D di x x ws= -1 # 1 1 This is not the only choice ... |

1 |
Alternative methods for CPS income imputation
- al
- 1986
(Show Context)
Citation Context ...en that MV x sj = ( ) i x x x d D si di si " =s= = 1 : } { + = D d dj sj x D x # 1 Other typical imputation algorithms include regression imputation [17,19], the EM algorithm, and hot-deck imputa=-=tion [7,14,4]-=-. Regression imputation selects donor cases in exactly the same way, and then calculates a least squares estimate: [ ] w x x si sjs= + 1 where [ ] dj D di x x ws= -1 # 1 1 This is not the only choice ... |

1 |
Boosting and Nave Bayesian Learning
- Elkan
- 1997
(Show Context)
Citation Context ... method differs from previous work in collaborative filtering [8] and imputation [18], in that we assume probabilities are conditionally independent. This assumption, which is also made in Nave Bayes =-=[5]-=-, enables us to pre-compute probabilities and store them in main memory, enabling very fast performance on millions of customers. The algorithm supports a variety of tunable parameters so that the met... |

1 |
System and method of predicting subjective reactions
- Hey
- 1989
(Show Context)
Citation Context ...RACT We develop a method for recommending products to customers with applications to both on-line and surface mail promotional offers. Our method differs from previous work in collaborative filtering =-=[8]-=- and imputation [18], in that we assume probabilities are conditionally independent. This assumption, which is also made in Nave Bayes [5], enables us to pre-compute probabilities and store them in ma... |

1 |
System and method for recommending items
- Hey
- 1991
(Show Context)
Citation Context ...le of revenues, percentages of spending, or page hits. We will use binary profiles in this article because this is what we have used in our experiments reported later. Collaborative filtering systems =-=[8,9,16]-=- implement a nearest neighbor variant of the above strategy. The donor set is restricted to the k closest matching customer profiles to a candidate. D = lowest(k) |x s1..N ,x d1..N | + = D d dj sj x k... |

1 |
The New Science of eMarketing: Targeted Email Communications to keep customers Coming Back
- Kaupp
- 2000
(Show Context)
Citation Context ...he store qtty returned as % of totalqty-0.330 percentage of products that the customer returns to the store 7. RELATED WORK Other researchers have reported similar results to those in our experiment. =-=[10]-=- reported a lift in clickthrough from 8.3% to 13.2% for market basket analysis (possibly similar to the method in this paper), and 13.96% for nearest neighbor method, in direct email campaigns (59% an... |

1 |
The Ecology of the Retail Store, Working paper, Vignette Corporation
- Kitts
(Show Context)
Citation Context ...elling products with high lift scores. In a past experiment we optimized shelf-layout by moving high lift items together. This resulted in a +40% increase in profit for items that were moved together =-=[11]-=-. The formula for lift is RecommendationValue(b) = Pr(b|a)/Pr(b) = Pr(a,b)/[Pr(a)*Pr(b)] Lift is a symmetric measure, so Lift(a,b)=Lift(b,a). A number greater than one is interpreted as the number of ... |

1 |
Automated collaborative filtering system
- Robinson
- 1998
(Show Context)
Citation Context ...le of revenues, percentages of spending, or page hits. We will use binary profiles in this article because this is what we have used in our experiments reported later. Collaborative filtering systems =-=[8,9,16]-=- implement a nearest neighbor variant of the above strategy. The donor set is restricted to the k closest matching customer profiles to a candidate. D = lowest(k) |x s1..N ,x d1..N | + = D d dj sj x k... |