A framework for promotion analysis in multi-dimensional space

Promotion is one of the most important elements in marketing. It is often desirable to find merit in an object (e.g., product, person, organization, or other business entity) and promote it in an appropriate community confidently. In this thesis, we motivate and discuss a novel class of data mining problems, called promotion analysis, for promoting a given object in a multi-dimensional space by leveraging object ranking information. The key observation is that most objects may not be highly ranked in the global space, where all objects are compared by all aspects; in contrast, there often exist interesting and meaningful local spaces in which the given object becomes prominent. Therefore, our general goal is to break down the data space and discover the most interesting local spaces in an effective and efficient way. We formally present the promotion analysis problem and formulate its variants and related notions. The promotion analysis problem is highly practical and useful in a wide spectrum of decision support applications. Typical application examples include merit discovery, product positioning and customer targeting, object profiling and summarization, identification of interesting features, and explorative search of objects. In fact, these applications are not new as they have been extensively studied and practiced in the marketing field. While existing commercial database and business intelligence systems can well support the functionality of retrieving the most highly ranked objects in some local space, there exists no multidimensional ranking analysis study for promotional purposes. Supporting effective and efficient online promotion analysis, nevertheless, presents many technical challenges, such as the spurious promotion problem, explosion of search space, and the high complexity of aggregation. Toward this end, we systematically study the problem and develop a general, principled promotion analysis framework. In terms of the search space, both subspaces formed on categorical dimensions and regions formed on continuous dimensions are examined. In terms of the object domain, both uniform object collection and multidimensional object space are studied. Moreover, we propose a unified query model to accommodate various scoring functions and redundancy-aware semantics. We also develop a statistical method to avoid spurious promotion results. For efficient query processing with a desirable balance between online and offline costs, we investigate exact algorithms as well as approximate algorithms with probabilistic guarantee. The promotion analysis framework not only provides an integrated solution for decision support applications, but also opens up new horizons for future research in other areas like information network analysis, text mining, and probabilistic data management.