A Markov Chain Substitution Scheme for …jblanche/presentations_slides/...A Markov Chain...
Transcript of A Markov Chain Substitution Scheme for …jblanche/presentations_slides/...A Markov Chain...
![Page 1: A Markov Chain Substitution Scheme for …jblanche/presentations_slides/...A Markov Chain Substitution Scheme for Approximation of Choice Models Jose Blanchet Columbia University Joint](https://reader031.fdocuments.net/reader031/viewer/2022040109/5e49b7828923776c36024d3a/html5/thumbnails/1.jpg)
A Markov Chain Substitution Scheme for
Approximation of Choice Models
Jose Blanchet
Columbia University
Joint work with
Guillermo Gallego and Vineet Goyal
![Page 2: A Markov Chain Substitution Scheme for …jblanche/presentations_slides/...A Markov Chain Substitution Scheme for Approximation of Choice Models Jose Blanchet Columbia University Joint](https://reader031.fdocuments.net/reader031/viewer/2022040109/5e49b7828923776c36024d3a/html5/thumbnails/2.jpg)
Assortment Planning Problem
2
1
2
n
S
r1
r2
rn
Choice Model
Assortment Problem
Find S to
How to Estimate?
How to Optimize?
![Page 3: A Markov Chain Substitution Scheme for …jblanche/presentations_slides/...A Markov Chain Substitution Scheme for Approximation of Choice Models Jose Blanchet Columbia University Joint](https://reader031.fdocuments.net/reader031/viewer/2022040109/5e49b7828923776c36024d3a/html5/thumbnails/3.jpg)
Choice Model
3
Ordered preference list (permutation) of items
Customer selects the most preferable item available
Most general choice model: distribution over all permutations
Tradeoffs
Complex choice model: hard to optimize
Simple choice model: not rich enough
![Page 4: A Markov Chain Substitution Scheme for …jblanche/presentations_slides/...A Markov Chain Substitution Scheme for Approximation of Choice Models Jose Blanchet Columbia University Joint](https://reader031.fdocuments.net/reader031/viewer/2022040109/5e49b7828923776c36024d3a/html5/thumbnails/4.jpg)
Common Choice Models
4
Multinomial Logit model: MNL (Plackett-Luce, MacFadden (1974))
utility parameters: wi for item i, attribute wi+Xi & Xi’s i.i.d. Gumbel.
choice probabilities (S+ = S U {0})
Easy to optimize (Talluri and van Ryzin (2004), Gallego et al. (2004))
Nested Logit model (Williams (1977), McFadden (1978))
Easy for some parameters of the model (Davis et al. (2012))
Mixture of Multinomial Logit Model
NP hard to optimize (Rusmevichientong et al. (2010))
PTAS for a constant number of mixtures
Model Selection: which is the right model?
True choice model is latent
We only observe choice or sales data
Error in model selection can lead to highly suboptimal decisions
![Page 5: A Markov Chain Substitution Scheme for …jblanche/presentations_slides/...A Markov Chain Substitution Scheme for Approximation of Choice Models Jose Blanchet Columbia University Joint](https://reader031.fdocuments.net/reader031/viewer/2022040109/5e49b7828923776c36024d3a/html5/thumbnails/5.jpg)
Related Work
5
Smith and Agrawal (2000), Netessine and Rudi (2003): two-step
dynamic substitution.
Other dynamic substitution & data inference: Saure and Zeevi
(2009), Rusmevichientong and Topaloglu (2009).
Farias et al. (2010)
Estimate distribution over permutation with sparsest support consistent data
Can efficiently provide estimator under some conditions
Vulcano, van Ryzin, Ratliff (2012)
EM algorithm to estimate a semi-parametric family of choice models
![Page 6: A Markov Chain Substitution Scheme for …jblanche/presentations_slides/...A Markov Chain Substitution Scheme for Approximation of Choice Models Jose Blanchet Columbia University Joint](https://reader031.fdocuments.net/reader031/viewer/2022040109/5e49b7828923776c36024d3a/html5/thumbnails/6.jpg)
This Talk
6
Markov chain based “Universal” choice “model” (really a
computational tool).
Can be estimated efficiently
O(n2) parameters
Universal approximation for all random utility models
Exact if the underlying model is MNL
Good approximation bounds for general random utility model
Efficient assortment optimization
![Page 7: A Markov Chain Substitution Scheme for …jblanche/presentations_slides/...A Markov Chain Substitution Scheme for Approximation of Choice Models Jose Blanchet Columbia University Joint](https://reader031.fdocuments.net/reader031/viewer/2022040109/5e49b7828923776c36024d3a/html5/thumbnails/7.jpg)
Markov Chain Based Model
7
New primitive for substitution behavior
i
0
1
n
If i is not available
No transitions out of state 0
Markovian model
After transition to state j, customer behaves like first choice being j
Specified by O(n2) transition probability parameters
No purchase alternative
![Page 8: A Markov Chain Substitution Scheme for …jblanche/presentations_slides/...A Markov Chain Substitution Scheme for Approximation of Choice Models Jose Blanchet Columbia University Joint](https://reader031.fdocuments.net/reader031/viewer/2022040109/5e49b7828923776c36024d3a/html5/thumbnails/8.jpg)
Estimating Markov Chain Model
8
i
0
1
n
If i is not available
No purchase alternative
Fraction of customers who select j
given the first choice is i
![Page 9: A Markov Chain Substitution Scheme for …jblanche/presentations_slides/...A Markov Chain Substitution Scheme for Approximation of Choice Models Jose Blanchet Columbia University Joint](https://reader031.fdocuments.net/reader031/viewer/2022040109/5e49b7828923776c36024d3a/html5/thumbnails/9.jpg)
Estimating Markov Chain Model
9
Suppose distribution over permutation σ given by p(σ)
![Page 10: A Markov Chain Substitution Scheme for …jblanche/presentations_slides/...A Markov Chain Substitution Scheme for Approximation of Choice Models Jose Blanchet Columbia University Joint](https://reader031.fdocuments.net/reader031/viewer/2022040109/5e49b7828923776c36024d3a/html5/thumbnails/10.jpg)
Estimating Markov Chain Model
10
Fraction of customers who select j
given the first choice is i
Data required to estimate the model
Choice probability data for n offer sets
Given S we can estimate
![Page 11: A Markov Chain Substitution Scheme for …jblanche/presentations_slides/...A Markov Chain Substitution Scheme for Approximation of Choice Models Jose Blanchet Columbia University Joint](https://reader031.fdocuments.net/reader031/viewer/2022040109/5e49b7828923776c36024d3a/html5/thumbnails/11.jpg)
Computing Choice Probability Estimates
11
i
0
j
k
S
i’
Define Markov chain for offer set S: M(S) All states in S (including 0) are absorbing states
Estimate of choice probability of item j in S
Can be computed efficiently for any j, S
No closed form expression
![Page 12: A Markov Chain Substitution Scheme for …jblanche/presentations_slides/...A Markov Chain Substitution Scheme for Approximation of Choice Models Jose Blanchet Columbia University Joint](https://reader031.fdocuments.net/reader031/viewer/2022040109/5e49b7828923776c36024d3a/html5/thumbnails/12.jpg)
Approximation Bounds: MNL Model
12
Suppose underlying model is MNL with parameters ui for all i
![Page 13: A Markov Chain Substitution Scheme for …jblanche/presentations_slides/...A Markov Chain Substitution Scheme for Approximation of Choice Models Jose Blanchet Columbia University Joint](https://reader031.fdocuments.net/reader031/viewer/2022040109/5e49b7828923776c36024d3a/html5/thumbnails/13.jpg)
Approximation Bounds: Other Models
13
McFadden and Train (2000)
Every random utility based model can be approximated
arbitrarily closely by a mixture of MNL
Suffices to prove approximation bounds for mixture models
Consider a mixture of MNL model with K segments Probability of segment k is
Parameters for segment k:
Assume wlog.
Choice Probability:
![Page 14: A Markov Chain Substitution Scheme for …jblanche/presentations_slides/...A Markov Chain Substitution Scheme for Approximation of Choice Models Jose Blanchet Columbia University Joint](https://reader031.fdocuments.net/reader031/viewer/2022040109/5e49b7828923776c36024d3a/html5/thumbnails/14.jpg)
Approximation Bounds: MMNL Model
14
MMNL model (with K segments)
For α=0.5, we get a 0.75-approximation for choice probabilities
![Page 15: A Markov Chain Substitution Scheme for …jblanche/presentations_slides/...A Markov Chain Substitution Scheme for Approximation of Choice Models Jose Blanchet Columbia University Joint](https://reader031.fdocuments.net/reader031/viewer/2022040109/5e49b7828923776c36024d3a/html5/thumbnails/15.jpg)
Approximation Bounds: MMNL Model
15
An example (2 classes of customers completely asymmetric utilities)
shows that bounds are sharp.
Numerical experiment with random uj’s & report average over 500
randomly picked offer sets S (of sizes 30% to 60% n).
Case n K=log(n) errMC(%)
1 10 3 3.1
2 20 3 2.4
3 30 4 2.5
4 40 4 2.4
5 60 5 1.9
6 80 5 1.6
7 100 5 1.6
8 150 6 1.2
9 200 6 1.1
10 500 7 0.8
11 1000 7 0.6
Average worst case relative error
in choice probabilities
![Page 16: A Markov Chain Substitution Scheme for …jblanche/presentations_slides/...A Markov Chain Substitution Scheme for Approximation of Choice Models Jose Blanchet Columbia University Joint](https://reader031.fdocuments.net/reader031/viewer/2022040109/5e49b7828923776c36024d3a/html5/thumbnails/16.jpg)
Approximation Bounds: MNL Model
16
Suppose underlying model is MNL with parameters ui for all i
![Page 17: A Markov Chain Substitution Scheme for …jblanche/presentations_slides/...A Markov Chain Substitution Scheme for Approximation of Choice Models Jose Blanchet Columbia University Joint](https://reader031.fdocuments.net/reader031/viewer/2022040109/5e49b7828923776c36024d3a/html5/thumbnails/17.jpg)
Approximation Bounds: Other Models
17
McFadden and Train (1996)
Every random utility based model can be approximated
arbitrarily closely by a mixture of MNL
Suffices to prove approximation bounds for mixture models
Consider a mixture of MNL model with K segments Probability of segment k is
Parameters for segment k:
Assume wlog.
Choice Probability:
![Page 18: A Markov Chain Substitution Scheme for …jblanche/presentations_slides/...A Markov Chain Substitution Scheme for Approximation of Choice Models Jose Blanchet Columbia University Joint](https://reader031.fdocuments.net/reader031/viewer/2022040109/5e49b7828923776c36024d3a/html5/thumbnails/18.jpg)
Approximation Bounds: MMNL Model
18
MMNL model (with K segments)
For α=0.5, we get a 0.75-approximation for choice probabilities
![Page 19: A Markov Chain Substitution Scheme for …jblanche/presentations_slides/...A Markov Chain Substitution Scheme for Approximation of Choice Models Jose Blanchet Columbia University Joint](https://reader031.fdocuments.net/reader031/viewer/2022040109/5e49b7828923776c36024d3a/html5/thumbnails/19.jpg)
Assortment Optimization
Optimization Problem
i
0
1
n
vs
![Page 20: A Markov Chain Substitution Scheme for …jblanche/presentations_slides/...A Markov Chain Substitution Scheme for Approximation of Choice Models Jose Blanchet Columbia University Joint](https://reader031.fdocuments.net/reader031/viewer/2022040109/5e49b7828923776c36024d3a/html5/thumbnails/20.jpg)
Conclusions
20
Choice model selection and assortment optimization problem
Present Markov chain based universal choice model
Simultaneous approximation for all random utility models (under mild
assumptions)
Polynomial time assortment optimization
Future directions
Additional constraints (eg. capacity) in assortment optimization
![Page 21: A Markov Chain Substitution Scheme for …jblanche/presentations_slides/...A Markov Chain Substitution Scheme for Approximation of Choice Models Jose Blanchet Columbia University Joint](https://reader031.fdocuments.net/reader031/viewer/2022040109/5e49b7828923776c36024d3a/html5/thumbnails/21.jpg)
21
Questions?