Researchers at Imperial College London Developed Deep Learning-Based Choice Models Under Feature-Free and Feature-Based Settings of Choice Modeling

Choice modeling tries to simulate a person’s or a group’s decision-making process using preferences disclosed or explicitly stated in a given setting or circumstances. In real-world situations, choice models support assortment, marketing, and price choices by assisting the company in better understanding customer behavior. It has been a critical subject in studying personal preference or utility in various disciplines, including operations research, marketing, operations management, and psychology.

Several methods exist in the literature, and we can distinguish two main approaches: Probabilistic choice models and Neural network models in utility modeling. The first category is also known as choice modeling under the feature-free setting. It targets to describe the likelihood of selecting a product from a selection (a set of offered products). The most famous technics belonging to feature-free based models are Maximum likelihood estimation (MLE), Expectation maximization (EM), and Linear programming (LP)-based method. The works in the second category often represent the latent utility of each product as a parameterized (neural network) function of the features. Then they use a multinomial logit model (MNL) to connect the choice probabilities with the latent utilities. These models learn primarily by recovering the function parameters from observations of consumer behavior in terms of choice. Recently, researchers from the University of London created choice models based on deep learning in feature-free and feature-based scenarios. The intrinsic value of each candidate’s choice and the impact of assortment on choice probability are captured by the suggested model.

The novel technique is the first deep learning-based choice approach to manage the feature-free setting and the first neural network model to explicitly incorporate the impact of the assortment on choice probabilities, enabling variable assortment size in the training data. Two deep networks are used: Gated-Assort-Net (GAsN), made by hidden fully-connected layers, and Res-Assort-Net (RAsN), which has two residual blocks. The authors proposed two different architectures for each network, one working in feature-free choice conditions and another for choice working with features. In the context of a feature-free choice problem, the proposed GAsN and RAsN architectures take the assortment vector as input and use it again to create an output gate to return the choice probability. However, for the two architectures of  GAsN and RAsN suggested for feature-based choice modeling, the input is formed by the latent utilities given by the feature encoder and the assortment vector. In addition, the authors propose to distinguish between customer features and product features. These two types of features are encoded to generate one latent utility for each product. Finally, the assortment is reused to ensure the choice probability. 

To validate the approach suggested in this article, the authors carried out a numerical experiment on two real datasets, SwissMetro and Expedia Search, which include dynamic customer/product attributes, which means that the features linked to each training sample may vary from the others. The obtained results proved that the two proposed neural networks perform better than the benchmark models. The outcome of this experiment demonstrates that, even when a product or customer feature is present, the assortment can still aid in better explaining customer preferences. The authors show that a single neural network model with a fixed design can recover both simple choice models like MNL and more sophisticated ones like  Markov chain choice mode and non-parametric choice model. Compared to the abovementioned approaches, the novel approach offers an efficient learning process with improved sample and computation efficiency. 

Through a binary representation of the assortment and feature encoders, the authors of this research developed a unified deep learning-based framework that applies to feature-free and feature-based settings. The new technique performs exceptionally well when the underlying model/training data is too complex to be explained by a simple model. Finally, an experimental study has proven that it exceeds the existing models in the state-of-the-art.

This Article is written as a research summary article by Marktechpost Staff based on the research paper 'Deep Learning for Choice Modeling'. All Credit For This Research Goes To Researchers on This Project. Check out the paper.

Please Don't Forget To Join Our ML Subreddit