Final Report Summary - ASSORTMENTCUSTOMIZE (Real-time assortment customization for products with dynamic content: theory and an application to interactive TV)
Recommendation systems are tools designed to recommend options based on other people's past choices (collaborative filtering), assuming prices are exogenously set. A typical example of a recommendation system is the one used by Amazon.com. These systems do not optimally learn how to assign products to people, so they might not learn that a consumer could be interested in products that are not popular yet. Thus, collaborative filtering systems are suboptimal because a more profitable product could only be shown if people were already buying it.
Another type of recommendation systems is the 'content filtering' system, which analyzes the content of a consumer's past purchases to then offer her a set of products similar to her own previous choices. Content filtering systems can also be sub-optimal because they would not offer profitable products that are too different from existing products.
The aim of this project is to develop a method that optimally customizes the assortment of products shown to users according to each user' individual characteristics. First, it identifies product category membership for each product using content analysis, and looks at consumer historical data to infer consumer's class. The actual complete list of consumer classes can be set a priori, or empirically derived via latent class analysis (typical classes are based on interest, purpose, situational circumstances, or life stage). Given product categories, consumer classes, product price and product cost, the optimization method then identifies the best shortlist of products to be offered to that consumer. Next, the consumer makes a decision (to choose or not a product from the list). The decision is then used to update the optimization algorithm.
The decision outcome is also stored as part of that consumer's history, and used to improve our estimates of consumer class in future opportunities. In short, the system will learn: Product category from product content, Consumer's class from history of past choices, the optimal matching of product categories to consumer classes from experimentation across consumers. In order to solve this optimization problem we model the matching of product categories to consumer classes as a multi-armed bandit which we will solve using dynamic programming. The computation of the solution of a dynamic program tends to take long time, but we can greatly reduce that to few seconds using table lookup based on tabling of the Gittins index. We expect that this method will help firms to mitigate the paradox of choice, as they will be offering customized sets of products (no more than 10 per customer at any given time) that meet consumer characteristics and are profit-maximizing. We will test it in one industry in particular (cable TV), but the theory will be general enough to be applied in other industries with dynamic content and individual observations.
During the first phase of the project the original model specification (based on Gittins index) was extensively redesigned. While the original specification allowed the model to optimally learn one component (preferences) it did not allow for the optimal learning of two components simultaneously (preferences and styles). The new specification allows for the simultaneous learning of any number of components. The fellow was able to secure a very large data set with choice data from a chain of video stores from Cambridge, MA. This data has a few million data points, and provides the empirical priors needed for the project. The fellow subsequently ran a very large amount of scenarios, re-wrote the optimization algorithm (now it is a faster, one-step look-ahead algorithm that allows for optimal learning) using R. These simulations gave the confidence to design and finish the next step: prototyping. The fellow subsequently finished the implementation of a web version of the ASSORTMENTCUSTOMIZE algorithm. Its database is running on MySQL. Its interface was coded in C#. The optimization engine was ported from R to C#. The questionnaire (to be used before the subjects reach the system) was built on Qualtrics.
The fellow then designed an experiment that will run in June 2013 with over 2000 respondents in a test-control condition to inspect the performance of our algorithm. During the write-up of the paper (late 2012) the fellow perfected the targeting and positioning. The fellow positions it as a recommendation systems paper, and decided to run the experiment with real-life users as described above. This way the paper will have a robust methodological contribution, as well as interesting data from a real-world context. The paper will be submitted to Management Science, an absolute top journal in the field.
Another type of recommendation systems is the 'content filtering' system, which analyzes the content of a consumer's past purchases to then offer her a set of products similar to her own previous choices. Content filtering systems can also be sub-optimal because they would not offer profitable products that are too different from existing products.
The aim of this project is to develop a method that optimally customizes the assortment of products shown to users according to each user' individual characteristics. First, it identifies product category membership for each product using content analysis, and looks at consumer historical data to infer consumer's class. The actual complete list of consumer classes can be set a priori, or empirically derived via latent class analysis (typical classes are based on interest, purpose, situational circumstances, or life stage). Given product categories, consumer classes, product price and product cost, the optimization method then identifies the best shortlist of products to be offered to that consumer. Next, the consumer makes a decision (to choose or not a product from the list). The decision is then used to update the optimization algorithm.
The decision outcome is also stored as part of that consumer's history, and used to improve our estimates of consumer class in future opportunities. In short, the system will learn: Product category from product content, Consumer's class from history of past choices, the optimal matching of product categories to consumer classes from experimentation across consumers. In order to solve this optimization problem we model the matching of product categories to consumer classes as a multi-armed bandit which we will solve using dynamic programming. The computation of the solution of a dynamic program tends to take long time, but we can greatly reduce that to few seconds using table lookup based on tabling of the Gittins index. We expect that this method will help firms to mitigate the paradox of choice, as they will be offering customized sets of products (no more than 10 per customer at any given time) that meet consumer characteristics and are profit-maximizing. We will test it in one industry in particular (cable TV), but the theory will be general enough to be applied in other industries with dynamic content and individual observations.
During the first phase of the project the original model specification (based on Gittins index) was extensively redesigned. While the original specification allowed the model to optimally learn one component (preferences) it did not allow for the optimal learning of two components simultaneously (preferences and styles). The new specification allows for the simultaneous learning of any number of components. The fellow was able to secure a very large data set with choice data from a chain of video stores from Cambridge, MA. This data has a few million data points, and provides the empirical priors needed for the project. The fellow subsequently ran a very large amount of scenarios, re-wrote the optimization algorithm (now it is a faster, one-step look-ahead algorithm that allows for optimal learning) using R. These simulations gave the confidence to design and finish the next step: prototyping. The fellow subsequently finished the implementation of a web version of the ASSORTMENTCUSTOMIZE algorithm. Its database is running on MySQL. Its interface was coded in C#. The optimization engine was ported from R to C#. The questionnaire (to be used before the subjects reach the system) was built on Qualtrics.
The fellow then designed an experiment that will run in June 2013 with over 2000 respondents in a test-control condition to inspect the performance of our algorithm. During the write-up of the paper (late 2012) the fellow perfected the targeting and positioning. The fellow positions it as a recommendation systems paper, and decided to run the experiment with real-life users as described above. This way the paper will have a robust methodological contribution, as well as interesting data from a real-world context. The paper will be submitted to Management Science, an absolute top journal in the field.