3. The parent-child or sibling relationship is established if the ratio of maximum overlap to the overall use of the tag t reaches the pre-defined threshold. The roles of the tags in relationship is determined as follows: if the tag to is used significantly more often than t, we declare t to be a child of to and vice versa If the usage of both tags is more or less equal, we declare them as sibling Before the assignment of relationships, we check whether it will not "break the context in the hierarchy, i. e, we compute an average overlap of all ances- tors or children in the branch respectively. If the average overlap falls below the threshold, we create a duplicate of tag to, assign tag t to it appropriately and make a new branch of it. The creation of the duplicate aims at solving the homonymy problem, where the to tag has multiple meanings, depending on a context Spreading Activation Spreading activation is a method for associative retrieval [7 in associative and semantic networks. hence a network data structure consisting of nodes and links modeling relationships between nodes. Searching is done by activating the se- lected node with an activation energy and spreading this energy through the edges to its neighbors so that Energylneighbor l- Energylorigin] The process runs recursively until convergence. At the end, the nodes activation levels represent the re of similarity to the initially chosen node or nodes. Spreading activation search in the folksonomy finds new relationships, which are not "visible when considering only set theory assumptions and the algo- rithm 1. If these relations are added to the already existing parent-child rela- tionships, it allows us to make contextual "jumps"between the tags, which we believe could have an interesting impact for the tag- based user modeling and Since the folksonomy does not provide direct links between tags, the spread- ng activation is performed on either the bipartite graph TI (tags-items)or TA (tags-actors)or on a combination of the two, all depending on the definition what neighbors means for the algorithm (i. e, the spreading activation on the Ta graph is performed if the neighbors of a tag are all actors which used that tag). However, due to the vast amount of connections present in every social tagging system, which leads to enormous time and space complexity of the re- cursive process, we needed to select a sub-graph of the whole folksonomy. For instance, for the TI bipartite graph, we select the sub-graph by modifying the neighbors function of a tag/item so that it will return only popular items/tags where popularity of a tag or item is defined(following the same principles as in algorithm 1)as a ratio of actors which used that particular tag or tagge that particular item. When spreading through TA graph, we define an actor as popular if he or she used at least k% of popular tags and tagged at least l% of popular items.3. The parent-child or sibling relationship is established if the ratio of maximum overlap to the overall use of the tag t reaches the pre-defined threshold. The roles of the tags in relationship is determined as follows: if the tag to is used significantly more often than t, we declare t to be a child of to and vice versa. If the usage of both tags is more or less equal, we declare them as siblings. 4. Before the assignment of relationships, we check whether it will not “break” the context in the hierarchy, i.e., we compute an average overlap of all ances￾tors or children in the branch respectively. If the average overlap falls below the threshold, we create a duplicate of tag to, assign tag t to it appropriately and make a new branch of it. The creation of the duplicate aims at solving the homonymy problem, where the to tag has multiple meanings, depending on a context. 2.2 Finding Related Tags by Spreading Activation Spreading activation is a method for associative retrieval [7] in associative and semantic networks, hence a network data structure consisting of nodes and links modeling relationships between nodes. Searching is done by activating the se￾lected node with an activation energy and spreading this energy through the edges to its neighbors so that Energy[neighbori ] = Energy[origin] |neighbors| The process runs recursively until convergence. At the end, the nodes’ activation levels represent the measure of similarity to the initially chosen node or nodes. Spreading activation search in the folksonomy finds new relationships, which are not “visible” when considering only set theory assumptions and the algo￾rithm 1. If these relations are added to the already existing parent-child rela￾tionships, it allows us to make contextual “jumps” between the tags, which we believe could have an interesting impact for the tag-based user modeling and personalization. Since the folksonomy does not provide direct links between tags, the spread￾ing activation is performed on either the bipartite graph T I (tags-items) or T A (tags-actors) or on a combination of the two, all depending on the definition what neighbors means for the algorithm (i.e., the spreading activation on the T A graph is performed if the neighbors of a tag are all actors which used that tag). However, due to the vast amount of connections present in every social tagging system, which leads to enormous time and space complexity of the re￾cursive process, we needed to select a sub-graph of the whole folksonomy. For instance, for the T I bipartite graph, we select the sub-graph by modifying the neighbors function of a tag/item so that it will return only popular items/tags, where popularity of a tag or item is defined (following the same principles as in algorithm 1) as a ratio of actors which used that particular tag or tagged that particular item. When spreading through T A graph, we define an actor as popular if he or she used at least k% of popular tags and tagged at least l% of popular items