Search

Graph Representation Learning aims to learn a rich and low-dimensional node embedding while preserving the graph properties. In this paper, we propose a novel Deep Attributed Graph Embedding (DAGE) that learns node representations based on both the topological structure and node attributes. DAGE a is able to capture, in a linear time and with a limited number of trainable parameters, the highly non-linear properties of attributed graphs. The proposed approach outperforms the current state-of-the-art approaches on node classification and node clustering tasks at a lower computational costs.

Providing personalized information to a wide audience, e.g., digital signage requires approaches and technologies different from those used for individual users. Such systems have to take care of user confidentiality and dynamically assess their interests. The paper developed earlier presented works on this topic and concentrates on the description of the ontological modeling of user interests and preferences as well as major components of such system. The developed ontological model intensively uses various types of relationships in order to provide for quantitative evaluation the user interest in a particular topic and semantic proximity of interests. It also enables generalization of interests, which is helpful for finding common interests for small amount of people, who are not related to each other and unlikely have much in common. Proposed structure of granular anonymous user profiles makes it possible to take care of the user confidentiality and to dynamically adapt to new information.

Concept information can be expressed by text, images or general objects which semantic meaning is clear to a human in a specific cultural context. For a computer, when available, text with its semantics (e.g., metadata, comments, captions) can convey more precise meaning than images or general objects with low-level features (e.g., color distribution, shapes, sound peaks) to extract the concept underlying the object. Among semantic measures, web-based proximity measures e.g., confidence, PMING, NGD, Jaccard, Dice, are particularly useful for concept evaluation, exploiting statistical data provided by search engines on terms and expressions provided in texts associated with the object.

This research includes the investigation, design and experimentation of models and measures of semantic and structural proximity for knowledge extraction and link prediction. The aim is to measure, predict and elicit, in particular, data from social or collaborative sources of heterogeneous information. The general idea is to use the information about entities (i.e. users) and relationships in collaborative or social repositories as an information source to infer the semantic context, and relations among the heterogeneous multimedia objects of any kind to extract the relevant structural knowledge. Contexts can then be used to narrow the domains and improve the performances of tasks such as disambiguation of entities, query expansion, emotion recognition and multimedia retrieval, just to mention a few.

Chapter 5 has laid out an approach to identify the semantic proximity between two named entities like brand and product names, locations, and so forth. The underlying chapter will now work on this model and put it to use for the identification and recommending of technology synergies. That is, given a set of technologies associated with different business units, we want to know which are the pairs of technologies that exhibit the largest synergies.

The two preceding chapters have demonstrated that classification taxonomies can be put to use in improving recommender systems in terms of the quality of their recommendations. The taxonomy we resorted to for all the empirical evaluations was the one from Amazon.com. Now we want to give an example how Web 2.0 taxonomies, having been crafted by collective efforts of several thousands of volunteering editors, can likewise be used to these ends.