Proactive Recommendation Based on $\mathcal{E}\mathcal{L}$ Concept Learning

Abstract

This paper proposes a novel knowledge-based approach to proactive recommendation. It exploits $\mathcal{E}\mathcal{L}$ concept learning to automatically model user preferences and is different from traditional knowledge-based approaches that require users to input explicit needs. Given an item being browsed, a set of marked items and an integer threshold k which is used to determine user preferences (where individuals are treated as items), the approach learns recommendatory restricted $\mathcal{E}\mathcal{L}$ concepts and returns unmarked instances of these concepts as recommendations. A recommendatory restricted $\mathcal{E}\mathcal{L}$ concept is a most specific restricted $\mathcal{E}\mathcal{L}$ concept that has at least k marked items, the item being browsed and at least one other unmarked item as instances. Intuitively, a recommendatory restricted $\mathcal{E}\mathcal{L}$ concept models a maximal set of user preferences. To guarantee that a learned concept has a finite size and the learning process is efficient, the proposed approach does not handle general $\mathcal{E}\mathcal{L}$ concepts but only restricted $\mathcal{E}\mathcal{L}$ concepts which have restrictions on the number of nested quantifiers and the inner occurrence of existential restrictions. This paper treats the problem of learning recommendatory restricted $\mathcal{E}\mathcal{L}$ concepts as the problem of maximal frequent itemset mining (MFI-mining) and presents an efficient MFI-mining algorithm to learn these concepts. Experimental results demonstrate the feasibility of the proposed approach in terms of efficiency and scalability.