Skip to content

Latest commit

 

History

History
28 lines (24 loc) · 2.83 KB

future-considerations.md

File metadata and controls

28 lines (24 loc) · 2.83 KB
Raw

Future considerations

  • Need to update based on comments

Current list

  • Wandering (serendipity) and forgetting probs and null models when setting either alpha or beta to 0
  • Heterogenity of alpha instead of setting one constant across all agents
  • Certain cost of learning, e.g. learning is easier for "closer" topics, and takes more time with high-degree topics (e.g. longer Wikipedia page)
  • Topic capacity in terms of information of the sub-knowledge tree
  • Consider the knowledge trees of individuals and analysis of such, i.e. remember the paths of learning/discovery in addition to only the topic nodes
  • Dependency of knowledge (directed) and strength of knowledge (weights; can consider forgetting and refinement to increase and decrease)
  • Learning through friends might need a primer, i.e. a certain topic to come up first connected to both agents, new topic needs to be neighbor of the common topic and the friend
  • Specialist and generalist specification and their distributions in different settings
  • Overlap of topics Js_T (Hamming similarity) of neighboring agents (or within certain distance) versus just random pairs of agents
  • Local diversity (consider topics of only neighboring agents or agents within certain distance) as a function of distance between agents
  • Nestedness of the bipartite graph
  • Persistent homology of the whole graph or the projected topic graph
  • Modularity (and number of modules) of the bipartite or projected topic graph
  • Growing and dynamic intralayer net with node/edge addition/removal, and edge reorganization to simulate changing friendship/human network and evolving knowledge graph
  • Consider multilayer nets, e.g. analysis in time as layers, or inclusion of more layers in addition to just agents and topics (need to come up with examples)
  • Relax some of the Heaviside step nonlinearity in the update equations and in update_via_matmul
  • Change the size of the topic and agent networks, e.g. (a) fix n_a but change n_t or (b) fix n_a + n_t but change their ratios
  • Data:
    • Use citation/paper networks, with authors as agents (assuming no agent groups), papers (or keywords) as topics, fields/subfields as groups of topics
    • Guessing alpha: assume intralayer networks as static but initialize interlayer network at a certain year, simulate with different alpha then calculate the likelihood of the evolved network to different actual years with matching density of the bipartite graph. This hopefully examines the tendency to self-learn or learn from friends at different times
    • Data as initial seed: use the empirical network as initialzed state, then simulate with different alpha and compare the different outcomes on topic diversity, modularity, ... Also compare with the results using the max-likelihood alpha from above (or projected alpha if MLE alpha seems to change with time)