The overall objective of this PhD is to develop learning and integration strategies of several modalities (visual,textual, structural, etc.) for an efficient classification of multimedia documents on the Web. The importance of this work have been highlighted by the proliferation of multimedia documents, in particular through the WEB. Paradoxically, despite this apparent abundance of data, the lack of data of sufficient quality and quantity in certain areas makes the task of automatic classification difficult. Therefore, this thesis work begins by proposing and applying methods of knowledge transfer where information on various categories, which are called source categories, can help the classification of target categories.