antosdk-apps/Blogger/build/debug/api/ai/cluster.lua

347 lines
9.3 KiB
Lua
Raw Permalink Normal View History

local doclassify = {}
local st = require("stmr")
doclassify.bow = function(data, stopwords)
-- first step get a table of worlds that contain
-- world: occurences
local bag = {}
for w in data:gmatch('%w+') do
local word = w:lower()
if not stopwords[word] then
word = st.stmr(word)
if bag[word] then
bag[word].count = bag[word].count + 1
else
bag[word] = {count=0, tf=0, tfidf=0.0}
bag[word].count = 1
end
end
end
-- now calculate the tf of the bag
for k,v in pairs(bag) do
bag[k].tf = math.log(1 + bag[k].count)
end
return bag
end
doclassify.len = function(table)
local cnt = 0
for k,v in pairs(table) do cnt = cnt+1 end
return cnt
end
doclassify.tfidf = function(documents)
-- now for each term in a bag, calculate
-- the inverse document frequency, which
-- is a measure of how much information
-- the word provides, that is, whether the
-- term is common or rare across all documents
local ndoc = doclassify.len(documents)
for k,bag in pairs(documents) do
-- for eacht term in bag
-- calculate its idf across all documents
for term,b in pairs(bag) do
local n = 0
for id,doc in pairs(documents) do
if doc[term] then n = n+1 end
end
--echo("term:"..term.." appears in"..n.." documents")
b.tfidf = b.tf*math.log(ndoc/n)
end
end
end
doclassify.search = function(term, documents)
local r = {}
for id, doc in pairs(documents) do
if doc[term:lower()] then
r[id] = doc[term].tfidf
end
end
return r
end
doclassify.get_vectors = function(documents)
-- get a list of vector from documents
local index = 0
local vectors = {}
local maps = {}
local terms = {}
local maxv = 0
for id in pairs(documents) do
maps[id] = {}
vectors[id] = {}
end
-- first loop, get the term
for id, doc in pairs(documents) do
for k,v in pairs(doc) do
-- get max value
if v.tfidf > maxv then
maxv = v.tfidf
end
-- get the term
if not terms[k] then
index = index + 1
terms[k] = index
end
for pid in pairs(documents) do
if not maps[pid][k] then
if id == pid then
maps[pid][k] = v.tfidf
else
maps[pid][k] = 0
end
else
if maps[pid][k] == 0 and id == pid then
maps[pid][k] = v.tfidf
end
end
end
end
end
-- reindexing the vectors
for id in pairs(documents) do
for k,v in pairs(maps[id]) do
vectors[id][terms[k]] = v
end
end
--echo("Max tfidf "..maxv.." in document #"..maxid.." of term "..term)
return vectors, maxv, index, terms
end
doclassify.similarity = function(va, vb)
-- using cosin similarity
local dotp = 0
local maga = 0
local magb = 0
for k = 1,#va do
dotp = dotp + va[k]*vb[k]
maga = maga + va[k]*va[k]
magb = magb + vb[k]*vb[k]
end
maga = math.sqrt(maga)
magb = math.sqrt(magb)
local d = 0
if maga ~= 0 and magb ~= 0 then
d = dotp/ (magb*maga)
end
return d
end
doclassify.similarities = function(v1, collection)
local similarities = {}
assert(#v1 == #(collection[1]), "Incorrect vectors size")
for i=1,#collection do
similarities[i] = doclassify.similarity(v1, collection[i])
end
return similarities
end
doclassify.mean_similarity = function(v1, v2)
assert(#v1 == #v2, "Incorrect vectors size")
local similarities = {}
for i = 1,#v1 do similarities[i] = doclassify.similarity(v1[i], v2[i]) end
return doclassify.mean(similarities)
end
doclassify.similarity_chart = function(id, vectors)
local vs = {}
local cnt = 0
local lut = {}
for k,v in pairs(vectors) do
if k ~= id then
cnt = cnt + 1
vs[cnt] = v
lut[cnt] = k
end
end
if not vs[1] then return {} end
return doclassify.similarities(vectors[id], vs), lut
end
doclassify.top_similarity = function(id, vectors, n, th)
local chart,lut = doclassify.similarity_chart(id,vectors)
--echo(JSON.encode(chart))
--echo(JSON.encode(lut))
if not lut or #lut <= 0 then return nil end
local top = {}
local j=0
local goon = true
if not th then
goon = false
th = 0
end
while j < n or goon
do
local i,maxv = doclassify.argmax(chart)
top[lut[i]] = maxv
chart[i] = 0.0
j=j+1
if maxv < th and goon then
goon = false
end
end
--for j=1,n do
-- local i,maxv = doclassify.argmax(chart)
-- top[lut[i]] = maxv
-- chart[i] = 0.0
--end
return top
end
doclassify.save_vectors = function(vectors, name)
local f = io.open(name,"w")
if f == nil then return false end
for id, v in pairs(vectors) do
f:write(id)
for i=1,#v do f:write(","..v[i]) end
f:write("\n")
end
f:close()
return true
end
doclassify.save_topchart = function(vectors, name,n)
local f = io.open(name,"w")
if f == nil then return false end
for k,v in pairs(vectors) do
local top = doclassify.top_similarity(k,vectors,n, 0.1)
for a,b in pairs(top) do
f:write(k.." "..a.." "..b.."\n")
end
end
f:close()
return true
end
doclassify.kmean = function(nclass, documents, maxstep, ids)
-- now
local vectors, maxv, size = doclassify.get_vectors(documents)
-- random centroids
local centroids = {}
local old_centroids = {}
local clusters = {}
--for pid in pairs(documents) do clusters[pid] = 0 end
-- add noise to mean_vector
for i = 1,nclass do
if ids == nil then
centroids[i] = doclassify.random(size,math.floor(maxv))
else
centroids[i] = vectors[ids[i]]
end
old_centroids[i] = doclassify.zeros(size)
end
-- loop until convergence or maxstep reached
local similarity = doclassify.mean_similarity(centroids, old_centroids)
local step = maxstep
while 1.0-similarity > 1e-9 and step > 0 do
clusters = {}
--echo(JSON.encode(centroids))
for id,v in pairs(vectors) do
local similarities = doclassify.similarities(v, centroids)
--echo(JSON.encode(similarities))
local cluster, maxvalue = doclassify.argmax(similarities)
--echo("doc #"..id.." is in clusters #"..cluster.." max value is "..maxvalue)
clusters[id] = cluster
end
-- storing the old centroids
old_centroids = centroids
-- calculate new centroids
local new_centroids = {}
for class in pairs(centroids) do
local cnt = 0
local cvectors = {}
for id,v in pairs(vectors) do
if clusters[id] == class then
cnt = cnt + 1
cvectors[cnt] = v
end
end
new_centroids[class] = doclassify.mean_vector(cvectors, size)
end
centroids = new_centroids
--echo(JSON.encode(centroids))
--echo(JSON.encode(old_centroids))
similarity = doclassify.mean_similarity(centroids, old_centroids)
echo("step #"..step..", similarity "..similarity)
step = step - 1
end
local results = {}
for i = 1,nclass do
local list = {}
local cnt = 0
for id,c in pairs(clusters) do
if c == i then
cnt = cnt + 1
list[cnt] = id
end
end
results[i] = list
end
return results, clusters, centroids
end
doclassify.zeros = function(n)
local vector = {}
for i = 1,n do vector[i] = 0.0 end
return vector
end
doclassify.random = function(n,maxv)
local vector = {}
for i=1,n do
vector[i] = math.random() + math.random(0, maxv)
end
return vector
end
doclassify.sum = function(v)
local sum = 0.0
for i=1,#v do sum = sum + v[i] end
return sum
end
doclassify.mean = function(v)
return doclassify.sum(v)/#v
end
doclassify.mean_vector = function(vectors, size)
local means = doclassify.zeros(size)
if not vectors or #vectors == 0 then return means end
--local size = #(vectors[1])
local times = 0
for k,v in pairs(vectors) do
for i=1,#v do means[i] = means[i] + v[i] end
times = times + 1
end
for i = 1,size do means[i] = means[i]/times end
return means
end
doclassify.argmin = function(v)
local minv = 0.0
local mini = 0.0
for i = 1,#v do
if v[i] <= minv then
mini = i
minv = v[i]
end
end
--echo("min index"..mini.." val "..minv)
return mini, minv
end
doclassify.argmax = function(v)
local maxv = 0.0
local maxi = 0.0
for i = 1,#v do
if v[i] >= maxv then
maxi = i
maxv = v[i]
end
end
return maxi,maxv
end
return doclassify