Business analytics has evolved from being a novelty used by a select few to an accepted facet of conducting business. Recommender systems form a critical component of the business analytics toolkit and, by enabling firms to effectively target customers with products and services, are helping alter the e-commerce landscape. A variety of methods exist for providing recommendations, with collaborative filtering, matrix factorization, and association-rule-based methods being the most common. In this paper, we propose a method to improve the quality of recommendations made using association rules. This is accomplished by combining rules when possible and stands apart from existing rule-combination methods in that it is strongly grounded in probability theory. Combining rules requires the identification of the best combination of rules from the many combinations that might exist, and we use a maximum-likelihood framework to compare alternative combinations. Because it is impractical to apply the maximum likelihood framework directly in real time, we show that this problem can equivalently be represented as a set partitioning problem by translating it into an information theoretic contextÑthe best solution corresponds to the set of rules that leads to the highest sum of mutual information associated with the rules. Through a variety of experiments that evaluate the quality of recommendations made using the proposed approach, we show that (i) a greedy heuristic used to solve the maximum likelihood estimation problem is very effective, providing results comparable to those from using the optimal set partitioning solution; (ii) the recommendations made by our approach are more accurate than those made by a variety of state-of-the-art benchmarks, including collaborative filtering and matrix factorization; and (iii) the recommendations can be made in a fraction of a second on a desktop computer, making it practical to use in real-world applications.
The sharing of databases either within or across organizations raises the possibility of unintentionally revealing sensitive relationships contained in them. Recent advances in data-mining technology have increased the chances of such disclosure. Consequently, firms that share their databases might choose to hide these sensitive relationships prior to sharing. Ideally, the approach used to hide relationships should be impervious to as many data-mining techniques as possible, while minimizing the resulting distortion to the database. This paper focuses on frequent item sets, the identification of which forms a critical initial step in a variety of data-mining tasks. It presents an optimal approach for hiding sensitive item sets, while keeping the number of modified transactions to a minimum. The approach is particularly attractive as it easily handles databases with millions of transactions. Results from extensive tests conducted on publicly available real data and data generated using IBM's synthetic data generator indicate that the approach presented is very effective, optimally solving problems involving millions of transactions in a few seconds.