escher/data/t.R

library(foreach)
library(iterators)
library(ggplot2)
library(dplyr)

#------

# messages <- read.csv2("chat/bus_time_data-1659191816218.csv", sep = ",")
# messages <- read.csv2("chat/zeta_molly_data-1659195953292.csv", sep = ",")

# messages <- messages %>% mutate(pub_time = as.POSIXct(strptime(pub_time, "%Y-%m-%d %T")))

messages <- read.csv2("transcript/minecraft_data-1659808309634.csv", sep = ",")
messages <- messages %>% mutate(
  start_time = as.POSIXct(strptime(start_time, "%Y-%m-%d %T")),
  end_time = as.POSIXct(strptime(end_time, "%Y-%m-%d %T"))
)

n_m <- nrow(messages)

#--------

ggplot(messages, aes(x=pub_time)) + theme_minimal() + geom_histogram(binwidth = 3600)

#--------

# Total length of messages
len <- as.numeric(messages[n_m, 1] - messages[1,1]) * 24 * 3600
# Poisson lambda for messages per second
lambda <- nrow(messages)/len

ggplot() + geom_histogram(mapping = aes(x=cumsum(rexp(n_m, lambda))), binwidth = 3600)

# Plot with overlaid posterior
ggplot(messages, aes(x=pub_time)) + theme_minimal() + 
  geom_histogram(binwidth = 600) + 
  geom_histogram(mapping = aes(x=cumsum(rexp(n_m, lambda)) + messages[1,1]), binwidth = 600, fill="red", alpha=0.3)

#-------

# Diffs and diff probabilities
diffps <- messages %>% transmute(diff = as.numeric(pub_time - lag(pub_time)), p = pexp(as.numeric(pub_time - lag(pub_time)), lambda))

ggplot(diffps, aes(x=diff)) + theme_minimal() +
  stat_ecdf()

# Diff with 0.01 quantile
qdiff <- qexp(0.01, lambda)

fmess <- messages %>% filter(
  as.numeric(difftime(pub_time, lag(pub_time), units = "secs")) < qdiff |
    as.numeric(difftime(lead(pub_time), pub_time, units = "secs")) < qdiff
)

#-------

# Print messages

for(i in 1:nrow(fmess)) {
  line = fmess[i,]
  
  cat(sprintf("[%s] <%s> %s\n", line$pub_time, line$nick, line$message))  
  
  if(difftime(fmess[i+1,"pub_time"], line$pub_time, units = "secs") > qdiff)
    cat("----------------------\n")
}

# Show where messages were selected
ggplot(messages, aes(x=pub_time)) + theme_minimal() + 
  geom_histogram(binwidth = 600) + 
  geom_histogram(data=fmess, binwidth = 600, fill="green", alpha=0.7)

#-------------

# Bayesian
# p(d < 3/60) = 0.01
# d ~ exp(l)
# l ~ gamma(a_p, b_p)
# l = argmax_l dgamma(a_p, 60)

a_p <- 13
b_p <- 60

# Total length of messages
# len <- as.numeric(difftime(messages[n_m,]$pub_time, messages[1,]$pub_time, units = "secs"))
len <- 749503.076
# Poisson lambda for messages per second

library(extraDistr)

gen_diffs <- rlomax(
  n_m, 
  1/(b_p + len),
  (a_p + n_m)
)

ggplot() + geom_histogram(mapping = aes(x=cumsum(gen_diffs)), binwidth = 3600)

# Plot with overlaid posterior
ggplot(messages, aes(x=pub_time)) + theme_minimal() + 
  geom_histogram(binwidth = 600) + 
  geom_histogram(mapping = aes(x=cumsum(rexp(n_m, lambda)) + messages[1,1]), binwidth = 600, fill="red", alpha=0.3)

#---------------

# Diff with 0.01 quantile
qdiff <- qlomax(0.01,   1/(b_p + len), (a_p + n_m))

fmess <- messages %>% filter(
  as.numeric(difftime(pub_time, lag(pub_time), units = "secs")) < qdiff |
    as.numeric(difftime(lead(pub_time), pub_time, units = "secs")) < qdiff
)