DeepBrowse: Similarity-Based Browsing Through Large Lists (Extended Abstract)

摘要

We propose a new approach for browsing through large lists in the absence of a predefined hierarchy. DeepBrowse is defined by the interaction of two fixed, globally-defined permutations on the space of objects: one ordering the items by similarity, the second based on magnitude or importance. We demonstrate this paradigm through our Wik-iBrowse app for discovering interesting Wikipedia pages, which enables the user to scan similar related entities and then increase depth once a region of interest has been found. Constructing good similarity orders of large collections of complex objects is a challenging task. Graph embeddings are assignments of vertices to points in space that reflect the structure of any underlying similarity or relatedness network. We propose the use of graph embeddings (DeepWalk) to provide the features to order items by similarity. The problem of ordering items in a list by similarity is naturally modeled by the Traveling Salesman Problem (TSP), which seeks the minimum-cost tour visiting the complete set of items. We introduce a new variant of TSP designed to more effectively order vertices so as to reflect longer-range similarity. We present interesting combinatorial and algorithmic properties of this formulation, and demonstrate that it works effectively to organize large product universes.