27-29 November, Vilnius
Conference about Big Data, High Load, Data Science, Machine Learning & AI
Conference is over. See you next year!
Amir Meimand is Zilliant Director of R&D, pricing scientist, where he designs and develops pricing solutions for customers and performs research in which he applies new methods to improve the current solutions as well as develop new tools. Prior to joining Zilliant, Amir helped design and develop a promotion planning and pricing platform for B2C retailers.
Amir holds a dual Ph.D. degree in Industrial Engineering and Operations Research from Pennsylvania State University. In his doctoral work, he applied operations research concepts to dynamic pricing and revenue management.
A Model-Based Recommendation Systems for B2B e-Commerce with Clustering and Association Rules Mining
In e-commerce world recommendation systems play a key role in elevating customer experience by reduce the number of clicks to find the items they want to purchase. In B2B e-commerce in addition to customer satisfaction another important aspect of recommendation system from business perspective is to increase the average order size over time. An effective recommendation system specially for B2B business should not only recommend items which were frequently purchased in the past by a customer but also should be able to identify the items which were never purchased in past but are likely to be in interest of the customers.
In this prestation we introduce a novel 2 steps method to design a recommendation system which can meet both goals. In the first step we employed association rules mining to compute the similarity between each pair of customers and in the second step a clustering method to identify the group of customers which are likely have similar purchasing behavior. We also present some analytical approach on how set and tune the hyper parameters of the clustering technique to get accurate result.
Bayesian Hierarchical Model for Predictive Analysis
Many predictive models require data to be structured (e.g., demographically, spatially, temporally). Each group of structured data has it is own parameters, but these parameters are related because data have a natural hierarchical structure. In such cases, two modeling approaches may be taken:
- All data may be modeled collectively;
2. Each group of data may be modeled separately.
With the first approach, there is a risk of ignoring both autocorrelation and latent differences, while with the second approach, there is a risk of ignoring latent similarity and data sparsity can be problematic. The best approach in such a situation is to develop a hierarchical model with the flexibility to capture and analyze the data structure, and the ability to account for and estimate effects from different groups. Hierarchical models are stronger because data can be analyzed across groups, thereby minimizing the effect of data sparsity. Higher level information can be shared effectively among the lower level groups, yet lower level estimation still follows its own data structure and pattern.
In this workshop we focus on the application of Bayesian hierarchical linear regression model in the area of pricing and revenue management. We will discuss how multi-layer model can be applied to hierarchal dataset to deal with data sparsity and reduce the noise to provide reliable and robust prediction. In this use case the behavior of every individual is modeled as a linear regression assuming each individual has its own unique behavior while there are some similarities in the behavior of the same groups.