Skip to content

Latest commit

 

History

History
318 lines (294 loc) · 20.8 KB

NEWS.md

File metadata and controls

318 lines (294 loc) · 20.8 KB

SimuNet 2.4.0

  • Removed used of |> and \(.) to bump down from R (>= 4.1) to R (>= 2.10)
  • Cleaned non-script files and folder

SimuNet 2.3.1

  • Releasing to Zenodo

SimuNet 2.3.0

  • Added scanList2igraph() and igraph2scanList() conversion functions:
    • does what their names implies
    • but also carries over the attrs attributes list
  • Added weightedAdj() to convert back a weightedAdj scanList to its 3D binary array form.
  • copy_attrs_to(): added an option to by-pass class modifications. If copy.class is set to FALSE, only attrs attribute is copied.

SimuNet 2.2.1

  • $ methods added for scanList to retrieve attributes in attrs:
    • Users can now use sL$name, e.g. sL$edge.Prob as a shortcut for attrs(sL,"edge.Prob").
  • [ methods added for scanList to subset scans:
    • sL[,,vec] is now also a scanList with attrs copied from sL.

SimuNet 2.2.0

  • customize_sampling() renamed into design_sampling(), with additional internal changes:
    • Now returns expDesign objects as the name suggests, and not just functions.
    • Removes the need to nest/chain design_exp() and customize_sampling(): now a single design_sampling() is enough, although design_exp(design_sampling()) would return an equivalent expDesign anyway.
  • design_exp() can now take combinations of previously created expDesign objects and functions, either named or anonymous (renamed into custom.FUN_X for now on):
    • NamedList() function added and used in design_exp() to keep track of function names input.
    • expDesign now prints the sequence of functions as a sequence of their names (more human-readable)

SimuNet 2.1.1

  • print method is cleaner for edgeProb and newly introduced edgeProbMat classes:
    • edgeProbMat is the matrix itself stored in edgeProb object (their $P), with attribute Beta prior

SimuNet 2.1.0

  • print methods are cleaner for SimuNet specific objects: *scanList objects printing now relies on Matrix sparseMatrix class
    • so do their "collapsed" weighted adjacency relative: weightedAdj (see below).
  • class sum (obtained via sum_scans()) has been renamed to weightedAdj:
    • Adj stored in attrs (and created by generate_edgeProb()) are now of class weightedAdj.
    • This change allows for homogeneous printing of weighted adjacency matrices
  • pkgdown is now used to update the website's package. Related files are in master branch's docs/ folder

SimuNet 2.0.2

  • enclosed asnr related example in \dontrun{} because some example were sometime not passing on unix systems

SimuNet 2.0.1

  • small edit in import_from_asnr()'s doc: now should pass R CMD check across systems (windows and unix)

SimuNet 2.0.0

  • Switch internal simulation framework to one relying on Bayesian inference:
    • edgeProb: edge presence probability matrices, renamed from presenceProb, are now drawn before each simulations from a posterior Beta distribution:
      • inferred from the observed weighted adjacency matrix and a prior Beta distribution (at the moment with alpha.prior = beta.prior = 0.5, i.e. Jeffrey's prior). Specifically, from a distribution Beta(Adj + alpha.prior,samp.effort - Adj + beta.prior)
      • edgeProb is drawn before each simulation: different edgeProbs will be obtained from a given Adj and samp.effort, but all of them come from this common posterior Beta distribution
      • This process of randomly drawing an edgeProb from a Beta distribution that would be used for all subsequent scans of a given simulation is akin to a Beta-Binomial process. However, "simply" drawing from a Beta-Binomial distribution would only provide a new weighted adjacency matrix, and not multiple binary scans that represent each scan, which is not desirable
      • Now, the desired behavior of having uncertainty and sampling effort included in the simulation should be attained
    • Uncertainty inherent from this Bayesian inference paradigm can be assessed via replicating many simulations, for which different edgeProb will be drawn
  • Switch from list of scans (list of 2D matrices) to 3 dimensional arrays:
    • the first 2 dimensions are the binary adjacency matrices for each scan
    • the 3rd dimension is the scan index
      This provides faster simulation and manipulations, but required an important rewriting of internal functions. Reliance on a new S3 class, scanList, should allow for fairly transparent manipulation, however
  • New scanList object are the new class of object returned by the new main wrapper, renamed simunet(). They now consist in 2 mains components:
    • the 3D array representing the edges at each scans, the actual object
    • a single attribute attrs, which is a list, containing all relevant attributes to keep track of
    • Operations/functions applied to scanList object should default to being applied to the 3D array. This approach (not having scanList object being primary a list) was used for intuitiveness
    • scanList attributes list attrs can be accessed via the new function attrs(), analogous to base's attr()
  • Empiricism and empirical scanLists have been rethought:
    • They now rely on a new class of objects and two main functions:
      • expDesign objects which consist in function sequences to apply to a theoretical scanList to obtain an empirical one. These functions can be viewed as experimental manipulations of interest to simulate, e.g. node removal, observation error, etc.
      • design_exp(): used to create expDesign objects. Inputted functions are chained via purrr's invoke() function.
      • perform_exp(): used to apply an expDesign objects to scanList objects
      • simunet() is now overall a wrapper to generate edgeProb, simulate a theoretical scanList, and can handle expDesign objects to directly obtain empirical scanLists. Otherwise, users can first generate a theoretical scanList and later apply expDesign to them via perform_exp()
    • In a similar fashion to data manipulation with tidyverse's functions, or similarly to how image editors can rely on layers, expDesign objects expect experimental manipulations to be chained and combined. Therefore, they should accept scanList objects as their first argument, and return a modified scanList onto which other manipulation could be applied. Special cases like collapsing binary scans into their sum within a weighted adjacency matrix (sum_scans()), or their scaled equivalent (scale_scans()), can also be used at the end of the function sequence.
    • Common manipulations are included in the package a experimental design "building blocks". Such include so far:
      • add_scans(): draws additional scans based on the inputted scanList's edge.Prob. For instance, to compensate expectable sampling effort difference between two experimental designs
      • remove_mostPeripheral(), to remove the overall most peripheral individual from all scans
      • sum_scans() and scale_scans(), to be use to collapse binary scans into weighted adjacency matrices, for instance at the end of a manipulation sequence
    • Empirical sampling, with group-scan and focal-scan sampling, are now handled as (part of) experimental designs passed as expDesign objects, as such "building blocks":
      • the function customize_sampling() can be used to define commonly defined group and focal sampling method, and can accept like previously:
        • in the case of group-scan sampling:
          • constant or random probabilities of observing an edge
          • observation probabilities derived from user-defined functions
        • in the case of focal-scan sampling:
          • even or random probabilities of observing an edge
          • focal selection probabilities derived from user-defined functions
      • customize_sampling() returns a function that can be used as a design "building block" within design_exp()
    • These "building blocks", as functions of scanLists returning scanLists, do not need to be used within expDesign object, but can verbosely chained/piped from scanList object (taking advantage of base's new pipe |> for instance). This is particularly useful for sum_scans() and scale_scans() for instance
    • Provided "building blocks" are also written to handle list of scanLists, i.e. list of class sLlist, and return a list of what they would have return. This allow for simple design of several expDesign objects passed simultaneously to be applied to a scanList object through perform_exp(), to later be able to compare their outputs.
  • Otherwise rework done on the whole code from the ground up, documentation improved with links, and generally attempt at optimization here and there

SimuNet 1.4.0

  • WIP: started writing tools to interact with the ANTs package. Barely started.

SimuNet 1.3.0.9000

  • Added a homemade class of packed matrices: snPackMat (SimuNet Packed Matrix) that stores only a vector of the relevant data, as well as a way to unpack the matrix to a regular one. Especially designed to avoid storing empry matrix triangles and performing useless operations on them. Now simu_scan has an optional use.snPackMat logical argument (FALSE by default). Early benchmarks are promising
  • Added two user-friendly functions to interact with the Animal Social Network Repository:
    • import_from_asnr can mostly be used with:
      • the Class and species folder (in the Networks folder of the github repository), and possibly graphml file name, as character strings
      • the URL of a graphml object within the repository
        an option allows the user to easily retrieve a igraph network object, or an adjacency matrix, from the graphml file
    • asnr_network_df retrieve the list (into a dataframe) of graphml file in the asnr repository. Internally used by import_from_asnr

SimuNet 1.2.0.9000

  • Switched internally to integer matrices instead of numeric matrices: significant improvement of -10% computation time and -30% memory allocation on a n = 21 total_scan = 9000 empirical network
  • Added dependency/reliance on the Matrix package, notably for its printing of sparse matrices

SimuNet 1.1.0

  • implemented a plot method for scan and empiScan objects. Internal code probably cleanable, but working for now: plot() can be used either on a scan/empiScan object (which will be passed through summary()) or on a summary.scan/summary.empiScan one. Notable plot arguments available: method to choose within c("both","theoretical","group","focal") for empiScan objects. Special case of the layout argument used to pass igraphs layout or layout_ functions internally. This way, a layout is determined before calling plot.igraph (wrapped in plot_emprical) in the case of method = "both", to ensure that all networks rely on the same layout to ease visual comparison.

SimuNet 1.0.1

  • fixed some ugly print outputs for long focalList objects

SimuNet 1.0.0

  • v1.0.0 Release of a fully functional and documented version of this network simulation framework, including a shift toward OOP internally

SimuNet 0.6.0.9000

  • Added track of X.scaled (X being "theoretical", "group", or "focal") for summary.scan and summary.empiScan objects

SimuNet 0.5.0.9000

  • Renamed more explicitely the X.scan components of scan and empiScan object, to make way for X.sum and X.sampled new matrices (cf. below).
  • The summarymethod used on scan and empiScan objects can now be used to create objects of new summary.scan and summary.empiScan objects, storing X.sum and X.sampled (X being "theoretical", "group", or "focal"), and with dedicated print methods
  • Adj and total_scan arguments are now optional when sampling.param is passed when using simu_scan
  • Added several required function for this purpose:
    • sum_scan.list is an equivalent of previous sum_up.scan(s), but cleaner. It sums up any list of scans counting NAs as zeros.
    • added resolve_NA, which is called before sum_scan.list in the case of empirical scans. Moved all NAs related functions to empirical_NA_tools.R
    • added sum_scan.sampled and count_NA functions to also count sampled (non-NA) edges

SimuNet 0.4.0.9000

  • Included scans.to.do variable (a scan index or a vector of 1:total_scan) in wrappers and nested functions to generate scan and empiScan objects. Now scans.to.do is stored in several internal and output objects
  • Vectorized wrappers and nested functions to produce lists of scans (raw.scan, theoretical.scan, group.scan or focal.scan empirical scans)
  • Updated several print methods to display vectorized or list components of scan, empiScan and samplingParam objects, truncate them when required to not overload output

SimuNet 0.3.0.9000

  • Added a wrapper to generate_samplingParam. User should interact with simu_samplingParam to create samplingParam objects for their simulations.
  • Vectorized the focal and scans.to.do (which now replace scan.number for more transparency) components of focal objects. Next version should also vectorize some components of samplingParam, obsProb, scan, and empiScan objects
  • Finalized a working version of simu_scan for single theoretical and empirical scan. Next version should allow the user to directly use simu_scan to generate a list of either types of scans (thus superseding former iterate_scans).
  • Homogenize some function and variable names.

SimuNet 0.2.0.9000

  • First shift toward integration of OOP elements into the simulation framework:
    • Revamp from the ground up of the do.scan part of the non-OOP previous version
    • Attempt at homogenizing SimuNet's internal syntax:
      • the wrappers - to perform the whole network simulations (former Boot_scan()), iterate single scans into weighted adjacency matrix (former iterate_scans()), and draw a single random binary scan (former do.scan()) - will now follow the syntax simu_*() (starting with the simu_scan() wrapper to supersede do.scan())
      • new simulation internal objects (cf. below) will have generator following the syntax generate_className(), and as needed print.className() and other relevant S3 class methods.
      • presently, their related code use variable names distinct from their className syntax, rather favoring a variable = class.name syntax internally
      • function names including internal ones are (hopefully) more explicit, and are now action verbs to follow coding "grammar" recommendations
  • For the OOP-transition purpose, creation of several simulation objects (S3 class):
    • presenceProb objects, generator and related S3 methods: calculate and store infos on the presence probability P of a tie at each scan for each dyad, from inputted adjacency matrix (Adj), sampling effort (total_scan), and igraph network mode
    • samplingParam objects, generator and related S3 methods: store all usefull parameters, some in the form of new object classes (S3), related to the empirical aspect of the network simulation. Presently, stores what's needed for a single binary focal scan, i.e. infos on a single focal (out of a list of focals). Specifically, samplingParam objects store:
      • method: a character scalar between "group","focal" and "both", indicating the chosen scan sampling method.
      • mode: a character scalar representing the igraph network mode.
      • obsProb objects (cf. below)
      • focal objects (cf. below)
    • obsProb objects, generator and related S3 methods: calculate and store a probability of observation P of an edge at each scan sampling for each dyad, from inputted user-defined function obs.prob_fun of (i,j,Adj) (that can also be used to pass a single [0,1] numeric to use as a constant, or the string "random" to have all dyad probability drawn from runif(n*n,0,1))
    • focalList objects, generator and related S3 methods: draw and store a list focals of total_scan focals (used internally as their indices, but also keep track of their names), from inputted user-defined function focal.prob_fun of (n,Adj) (that should return a vector of n probability for each node to be drawn at each scan, but that can also be used to pass as the strings "random" or "even" to have them drawn from sample() or to maximize the evenness of drawn focals in the focal list)
    • focal objects, generator and related S3 methods (including plot.scan S3 method to plot the network relying on plot.igraph): determine and store which focal to sample, from inputted focalList object (cf. above) and scan.number (out of the sampling effort total_scan)
    • scan objects, generator and related S3 methods: draw and store, from inputted presenceProb object:
      • a raw scan: directed binary matrix drawn from the probability of presence of edges contained in a presenceProb object
      • a theoretical scan: the raw scan to which the selected mode has been applied
      • a scan.type scan string: for scan object not sampled from yet (i.e. not empirical scans of the empiScan class (cf.below)), scan$scan.type = "theoretical", but later in the process can become "empirical"
      • the original adjacency matrix Adj, sampling effort total_scan, selected igraph network mode, the probability matrix presence.prob (from presenceProb$P), and other parameters mostly for internal use
    • empiScan objects, generator and related S3 methods: draw and store, from inputted scan and samplingParam objects:
      • inherits from the scan S3 class
      • but have their scan.type = "empirical"
      • a method string from samplingParam$method
      • and both/either a group and/or focal scan: matrix/matrices containing* 0,1, or NA, which represents an edge that hasn't been sampled. Internally sample from scan$theoretical, to which the igraph network mode was already applied, and is set to keep NAs where they were drawn (relevant for "group" scan sampling), even if this result in a undirected adjacency matrix where NAs are non-symetrical. A function to minimize solvable NAs will be soon introduced: for mode = "max", this is e.g. when scan$raw[i,j] = 1 and scan$raw[j,i] = NA or inversly, for which both values can be set to 1; for mode = "min", this is e.g. when scan$raw[i,j] = 0 and scan$raw[j,i] = NA or inversly, for which both values can be set to 0. *: with mode = "plus", possible values also include 2
  • Revamping from the ground up as the new object were formalized and introduced in the algorithm, the internal code also got cleaned and fixed in some places as code was recycled from its previous non-OOP state:
    • Some functions were updated to split some of their overall work into additional more explicit and more "single-step/purpose" functions (e.g. generate_obsProb now relies on determine_obs.prob_type and calculate_obs.prob instead of containing all the code in itself)
    • empiScan$focal now show not only the line of the focal sampled, but also its column
    • rethinking of the way the igraph network mode is applied: apply_mode now keep track of an always "directed" raw scan, makes it symmetrical into a theoretical scan if an undirected mode is selected ("undirected", "max", "min", or "plus"), from which a focal and group scans are sampled through sample_from_scan (that internally relies on group_sample and focal_sample). Through this changes, internal algorithm changed significantly, especially in how NAs are handled in group scans. Also, zero_NA has been generalized into replace_NA, with subsequent changes in compare_with_transposed, but these are not used anymore within apply_mode and are kept for now just in case.
    • removed normally-superseded *.R files (focal.list.R, obs.prob_._tools), renamed do.scan.R into do.scan.old.R

SimuNet 0.1.0

  • Imported simulation-oriented functions from ConfiNet
  • Structure package till it checks ok on its own
  • Commented and removed mention of decide_use.rare.opti() for now till this routine is cleaned. Updated examples in Boot_scans() and bootstrap tools to not call for this function through setting use.rare.opti = FALSE
  • Added a NEWS.md file to track changes to the package.