Hidden layer models for company representations and product recommendations

Abstract

An increasing amount of marketing intelligence data is becoming available today. This includes data that describes information technology (IT) inventories, i.e. IT products purchased by companies. It is advantageous for hardware and software service providers to analyze this data and build recommender systems to propose new products for client companies. Real-time recommendations are usually done based on matrix factorization methods or association rules. In this work we study the applicability of generative models to the recommendation problem. We focus on models that are able to reveal latent connections between companies and deployed IT products and build discriminative features of the IT structure of a company. Additionally generative models of company-product data are of interest for service providers for efficient company comparison, application of similar marketing strategies towards the groups of similar companies. In this work, we first formalize the notion of a company and its IT install base. Then, we estimate various generative models that are able to reveal hidden structures in data. These are mainly topic and language modeling techniques emerging in natural language processing to the task of company-product modeling and sequential models that are widely used for product recommendations. More precisely, the analysis is done using (a) Latent Dirichlet Allocation (LDA) with the products in a company are treated as a set, (b) n-gram models or sequential association rules and (c) Recurrent Neural Networks (RNN), when the time of product appearance is taken into account. The techniques are used for a corpus consisting of 860k companies. The results of the study demonstrate that simpler generative models with lower number of parameters, such as LDA, fit company-product data better and are more beneficial for company IT install base modeling both in terms of goodness of fit of the model and recommendation quality.