As a technique to investigate behaviors of a computer network with low operational cost, network tomography has received considerable attentions in recent years. Most studies assume that the topology of the network of interest is known, and try to propose computationally and/or statistically efficient methods to estimate link-level properties, delay distribution, band- width etc., or global traffic properties such as point-to-point traffic matrix. Little progresses have been made for scenarios when topology of the target network is unknown, although it is often the case in many practical applications. The few published works on topology tomography entangle the problem primarily by clustering analysis, and usually work for binary trees only and often suffer from stable performance. In this article, we propose a new perspective to resolve the problem. By connecting the problem of topology tomography to the classic machine learning problem of “market basket analysis,” we find that simultaneous topology and loss tomography can be achieved by discovering association patterns of loss records collected at receivers, which can be efficiently resolved with light modifications of a recently developed statistical method known as the “theme dictionary model.” Both theoretical analysis and simulation studies prove the effectiveness of the novel approach for networks of tree as well as general topology.