Choice Modeling and Assortment Optimization on the Transformer Model
The problem of modeling customer choices and finding assortments with maximal revenue has been widely studied in revenue management. Random utility models (RUMs) are typically used to model choice. These models implicitly enforce a rational decision making process whereby a customer is endowed with...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Published: |
Massachusetts Institute of Technology
2024
|
| Online Access: | https://hdl.handle.net/1721.1/153714 https://orcid.org/0009-0004-2183-5338 |
| _version_ | 1811095037119823872 |
|---|---|
| author | Jiang, Qingxuan |
| author2 | Levi, Retsef |
| author_facet | Levi, Retsef Jiang, Qingxuan |
| author_sort | Jiang, Qingxuan |
| collection | MIT |
| description | The problem of modeling customer choices and finding assortments with maximal revenue has been widely studied in revenue management. Random utility models (RUMs) are typically used to model choice. These models implicitly enforce a rational decision making process whereby a customer is endowed with utilities for each product in the assortment and picks the product that maximizes her utility. This work seeks to explore a general class of choice models where the customer’s decision making process is not constrained in this fashion.
To allow for departures from rational choice (and RUMs), we posit that the customer indirect utility associated with a product is a function of the assortment offered to her. Motivated by the success of transformer models in deep learning, we investigate the case where this utility function is defined through a trained transformer network. This leads to a new class of neural network-based discrete choice models, which we call transformer choice models. The universal approximation property of the transformer network ensures that our model can approximate any discrete choice model, and thus it can capture irrationalities in choice behavior.
We perform computational experiments with real data to verify the generalization performance of our transformer choice model to unseen assortments. To ensure that our model does not overfit on the training data, we use dropout as the regularization method during training. We compare our model to both traditional choice models (the multinomial logit model and its synergistic variant that considers cross-product interaction) and machine learning-based choice models (decision forest choice model and feedforward neural network choice model) on two datasets: a large grocery panel dataset and an online hotel search dataset. We show that on both datasets, the transformer choice model has generalized well to unseen assortments with proper regularization. Moreover, on the more complex dataset of online hotel search, the transformer choice model has outperformed all other models in terms of out-of sample error.
We finally consider the assortment optimization problem on transformer choice models. While the general assortment optimization problem is complex and in-tractable, we empirically evaluate and compare several heuristic algorithms, including random search, quadratic approximation, and local search. Our experiments on transformer choice models with real prices show that a simple local search heuristic finds the global optimum for the assortment optimization problem in three-fourths of the data categories, while achieving a good approximation on the rest of the categories. This shows that in practice, local search can be a reasonable heuristic for assortment optimization on transformer choice models. |
| first_indexed | 2024-09-23T16:09:30Z |
| format | Thesis |
| id | mit-1721.1/153714 |
| institution | Massachusetts Institute of Technology |
| last_indexed | 2024-09-23T16:09:30Z |
| publishDate | 2024 |
| publisher | Massachusetts Institute of Technology |
| record_format | dspace |
| spelling | mit-1721.1/1537142024-03-14T03:05:01Z Choice Modeling and Assortment Optimization on the Transformer Model Jiang, Qingxuan Levi, Retsef Farias, Vivek Massachusetts Institute of Technology. Operations Research Center The problem of modeling customer choices and finding assortments with maximal revenue has been widely studied in revenue management. Random utility models (RUMs) are typically used to model choice. These models implicitly enforce a rational decision making process whereby a customer is endowed with utilities for each product in the assortment and picks the product that maximizes her utility. This work seeks to explore a general class of choice models where the customer’s decision making process is not constrained in this fashion. To allow for departures from rational choice (and RUMs), we posit that the customer indirect utility associated with a product is a function of the assortment offered to her. Motivated by the success of transformer models in deep learning, we investigate the case where this utility function is defined through a trained transformer network. This leads to a new class of neural network-based discrete choice models, which we call transformer choice models. The universal approximation property of the transformer network ensures that our model can approximate any discrete choice model, and thus it can capture irrationalities in choice behavior. We perform computational experiments with real data to verify the generalization performance of our transformer choice model to unseen assortments. To ensure that our model does not overfit on the training data, we use dropout as the regularization method during training. We compare our model to both traditional choice models (the multinomial logit model and its synergistic variant that considers cross-product interaction) and machine learning-based choice models (decision forest choice model and feedforward neural network choice model) on two datasets: a large grocery panel dataset and an online hotel search dataset. We show that on both datasets, the transformer choice model has generalized well to unseen assortments with proper regularization. Moreover, on the more complex dataset of online hotel search, the transformer choice model has outperformed all other models in terms of out-of sample error. We finally consider the assortment optimization problem on transformer choice models. While the general assortment optimization problem is complex and in-tractable, we empirically evaluate and compare several heuristic algorithms, including random search, quadratic approximation, and local search. Our experiments on transformer choice models with real prices show that a simple local search heuristic finds the global optimum for the assortment optimization problem in three-fourths of the data categories, while achieving a good approximation on the rest of the categories. This shows that in practice, local search can be a reasonable heuristic for assortment optimization on transformer choice models. S.M. 2024-03-13T13:28:54Z 2024-03-13T13:28:54Z 2024-02 2024-02-12T01:41:39.383Z Thesis https://hdl.handle.net/1721.1/153714 https://orcid.org/0009-0004-2183-5338 In Copyright - Educational Use Permitted Copyright retained by author(s) https://rightsstatements.org/page/InC-EDU/1.0/ application/pdf Massachusetts Institute of Technology |
| spellingShingle | Jiang, Qingxuan Choice Modeling and Assortment Optimization on the Transformer Model |
| title | Choice Modeling and Assortment Optimization on the Transformer Model |
| title_full | Choice Modeling and Assortment Optimization on the Transformer Model |
| title_fullStr | Choice Modeling and Assortment Optimization on the Transformer Model |
| title_full_unstemmed | Choice Modeling and Assortment Optimization on the Transformer Model |
| title_short | Choice Modeling and Assortment Optimization on the Transformer Model |
| title_sort | choice modeling and assortment optimization on the transformer model |
| url | https://hdl.handle.net/1721.1/153714 https://orcid.org/0009-0004-2183-5338 |
| work_keys_str_mv | AT jiangqingxuan choicemodelingandassortmentoptimizationonthetransformermodel |