Product Comparison Matrices (PCMs) are widely used for documenting or comparing a set of products. They provide a compact representation of the characteristics or features of the products in the form of a product by feature matrix. The expected benefits of such simple format are the ease of use and definition. However, there exist no standard, guidelines or specific tool support for creating PCMs. The various domains addressed in PCMs and the numerous contributors leads to a large diversity in the structure and content of these matrices. This creates a syntactic and semantic gap between the concrete representation of a PCM and its human interpretation. To address this problem, the objective of this thesis is to develop techniques for formalizing PCMs. We explore two possible solutions to reach our objective. Our first solution is to directly model PCMs through the use of a model-based approach. The central element of our approach is a domain metamodel that defines the structure and semantics of PCMs. It is completed by a set of transformations for automatically formalizing existing PCMs into PCM models. In our context, the elaboration of a model-based approach is a challenging task. The lack of standard or oracle hinders the evaluation of our approach. The presence of potentially millions of PCMs on the Web makes the manual building of an oracle unrealistic. To address these challenges, we propose an iterative process driven by data, end-users and services. It helps to evaluate our approach on a large number of PCMs while guiding the selection of a few PCMs to manually analyze. To ease the development of services on top of PCM models, we also create task-specific metamodels in relation to the edition, manipulation, import and export of PCMs. Our second solution is the synthesis of Feature Models (FMs) from PCMs. Similarly to PCMs, FMs represent product lines which are sets of products sharing commonalities and presenting variabilities. FMs are the most widespread formalism for modeling product lines. They allow to develop reasoning capabilities and numerous applications such as generating comparators or configurators which can be beneficial for PCMs. Yet FMs provide a different view on a product line than PCMs. PCMs also contain types of data that cannot be represented by FMs. To address this problem, we consider as input of our algorithm, a class of PCMs that we call configuration matrices. Such matrices present a simple structure and contain data that can be represented by FMs. We also target an extension of FMs called Attributed FMs (AFMs) as output of our algorithm. AFMs allow the modeling of common types of data in PCMs such as booleans, numbers and enumerations. The analysis of the state of the art reveals that there exist no algorithm for synthesizing AFMs. Moreover, existing synthesis techniques often foster precision over quality of the resulting FM or vice versa. This hinders the exploitation of the generated FM. We improve state of the art of FM synthesis with techniques that produce FMs that both exactly represent the information contained in the input and exhibit a meaningful hierarchy. Then, we extend this technique to produce AFMs from configuration matrices. These improvements contribute to allow a qualitative and precise formalization and exploitation of a class of PCMs through the use of AFMs. With our contributions, we provide two generic, extensible and complementary approaches for the formalization and exploitation of PCMs. This is a first step towards the creation of a community of users and contributors of PCMs. It opens new perspectives for the development of innovative services on top of PCMs coming from various sources.