Analysis of log data from Multi-Touch-Table at Herzog-Anton-Ulrich-Museum (HAUM)
Go to file
2023-09-28 15:04:59 +02:00
code Worked on second presentation, added problems and plots, figs, etc. 2023-09-28 15:04:59 +02:00
README.Rmd Started working on second presentation; created plots for it 2023-09-26 18:39:02 +02:00

---
title: "Background information about MTT data"
author: "Nora Wickelmaier"
date: "`r Sys.Date()`"
output: 
  html_document:
    number_sections: true
    toc: true
---

```{r, include = FALSE}
# setwd("C:/Users/nwickelmaier/Nextcloud/Documents/MDS/2023ss/60100_master_thesis")
devtools::load_all("../../../software/mtt")
```

# Log data from the Multi-Touch Table at the HAUM

The Multi Touch Table at the Herzog-Anton-Ulrich-Museum (HAUM) in
Braunschweig gives visitors of the Museum the opportunity to interact with
67 artworks and 3 tiles containing information about the museum and its
layout. The table was installed at the institute in October 2016 and since
November 2016 log files from interactions of visitors of the museum have
been collected. These log files are in an unstructured format and cannot be
easily analyzed. The purpose of the following document is to describe how
the data haven been transformed and which decisions have been made a long
the way.

# Data structure

The log files contain lines that indicate the beginning and end of possible
actions that can be performed when interacting with the artworks on the
table. The layout of the table looks like 70 pictures have been tossed on a
large table. Every artwork is visible at the start configuration. People
can move the pictures on the table, they can be scaled and rotated.
Additionally, the virtual picture cards can be flipped in order to find
more information of the artwork on the "back" of the card. One has to press
a little `i` for more information in one of the bottom corners of the card.
On the back of the card two (?) to six information cards can be found with
a teaser text about a certain topic. These topic cards can be opened and a
hypertext with detailed information pops up. Within these hypertexts
certain technical terms can be clicked for lay people to get more
information. This also opens up a pop-up. The events encoded in the raw log
files therefore have the following structure.

```
"Start Application"     --> Start Application
"Show Application"
"Transform start"       --> Move
"Transform stop"
"Show Info"             --> Flip Card
"Show Front"
"Artwork/OpenCard"      --> Open Topic
"Artwork/CloseCard"
"ShowPopup"             --> Open Popup
"HidePopup"
```

The right side shows what events can be extracted from these raw lines. The
"Start Application" is not an event in the original sense since it only
indicates if the table was started or maybe reset itself. This is not an
interaction with the table and therefore not interesting in itself. All
"Start Application" and "Show Application" are therefore excluded from the
data when further processed and are only in the raw log files.

# Parsing the raw log files

The first step is to parse the raw log files that are stored by the
application as text files in a rather unstructured format to a format that
is better handled. The data are therefore transferred to a spread sheet
format. The following section describes what problems were encountered
while doing this.

## Corrupt lines

When reading the files containing the raw logs into R, a warning appears
that says

```
Warning messages:
  incomplete final line found on '_2016/2016_11_18-11_31_0.log'
  incomplete final line found on '_2016/2016_11_18-11_38_30.log'
  incomplete final line found on '_2016/2016_11_18-11_40_36.log'
  ...
```

When you open these files, it looks like the last line contains some binary
content. It is unclear why and how this happens. So when reading the data,
these lines were removed. A warning will be given that indicates how many
files have been affected.

## Units of the variables

* Welche Einheit haben x und y? Pixel? --> yes
* Welche Einheit hat scale? --> some kind if bit, does not matter, when
  calculating a ratio
* rotation wirklich degree? --> yes
* Nach welchem Zeitintervall resettet sich der Tisch wieder in die
  Ausgangskonfiguration? --> PM needs to look it up

## How unclosed events are handled

## How a case is defined

* Herausfinden, ob mehr als eine Person am Tisch steht?
  - Sliding window, in der Anzahl von Artworks gezählt wird? Oder wie weit
    angefasste Artworks voneinander entfernt sind?
  - Man kann sowas schon "sehen" in den Logs - aber wie kann ich es
    automatisiert rausziehen? Was ist meine Definition von
    "Interaktionsboost"?
  - Egal wie wir es machen, geht es auf den "Event-Log-Daten"?

## Additional meta data

* Anreicherung der Log-Daten mit weiteren Metadaten? Was wäre interessant?

  - Metadata on artworks like, name, artist, type of artwork, epoch, etc.
  - School vacations and holidays
  - Special exhibits at the museum
  - Number of visitors per day (bei Sven noch mal nachhaken?)
  - Age structure of visitors per day?
  - ... ????

# Problems and how I handled them

This lists some problems with the log data that required decisions. These
decisions influence the outcome and maybe even the data quality. Hence, I
tried to document how I handled these problems and explain the decisions I
made.

## Weird behavior of `timeMs` and neg. `duration` values

I think the negative duration values happen, when an event starts in one
log file and completes in another one. The variable `timeMs` seems to be
continuous within one log file but not over several log files.


```{r, results = FALSE, fig.show = TRUE}
# Read data
dat0 <- read.table("data/haum/raw_logfiles_small_2023-09-26_13-50-20.csv", sep = ";",
                   header = TRUE)
dat0$date <- as.POSIXct(dat0$date)
dat0$glossar <- ifelse(dat0$artwork == "glossar", 1, 0)

# Remove irrelevant events
dat <- subset(dat0, !(dat0$event %in% c("Start Application",
                                        "Show Application")))

# Add trace variable
artworks <- unique(stats::na.omit(dat$artwork))
artworks <- artworks[artworks != "glossar"]
glossar_files <- unique(subset(dat, dat$artwork == "glossar")$popup)
glossar_dict <- create_glossardict(artworks, glossar_files,
                    xmlpath = "data/haum/ContentEyevisit/eyevisit_cards_light/")
dat1 <- add_trace(dat, glossar_dict)

# Close events
dat2 <- rbind(close_events(dat1, "move"),
              close_events(dat1, "flipCard"),
              close_events(dat1, "openTopic"),
              close_events(dat1, "openPopup"))
dat2 <- dat2[order(dat2$date.start, dat2$fileId), ]

plot(timeMs ~ as.factor(fileId), dat[1:5000,], xlab = "fileId")
```

The boxplot shows that we have a continuous range of values within one log
file but that `timeMs` does not increase over log files. Since it seems not
possible to fix this in a consistent way, I set all durations to `NA` where
`fileId.start` and `fileId.stop` are not identical. I kept `timeMs.start`
and `timeMs.stop` and also `fileId.start` and `fileId.stop` in the data
frame, so it is clear why there are no durations. The other

NOTE: Part of this problem was that time stamps that are part of the log
file names are not zero-left-padded and therefore the files were not in the
correct order when read into R. When zero left padding these file IDs and
sorting by them and then by `date.start` within, some of the durations are
exactly fixed. Unfortunately, only three `move` events were fixed, since it
only fixed irregularities *within* one log file. See below for more
details.

UPDATE: By now I remove all events that span more than one log file. This
lets me improve speed considerably.

## Left padding of file IDs

The file names of the raw log files are automatically generated and contain
a time stamp. This time stamp is not well formed. First, it contains an
incorrect month. The months go from 0 to 11 which means, that the file name
`2016_11_15-12_12_57.log` was collected on December 15, 2016 at 12:12 pm.
Another problem is that the file names are not zero left padded, e.g.,
`2016_11_15-12_2_57.log`. This file was collected on December 15, 2016 at
12:02 pm and therefore before the file above. But most sorting algorithms,
will sort these files in the order shown below. In order to preprocess the
data and close events that belong together, the data need to be sorted by
events and artworks repeatedly. In order to get them back in the correct
time order, it is necessary to order them based on three variables:
`fileId`, `date.start` and `timeMs`. The file IDs therefore need to
sort in the correct order (again see below for example). I zero left padded
the log file names within the data frame using it as an identifier. These
"file names" do not correspond exactly to the original raw log file names.
This needs to be kept in mind when doing any kind of matching etc.

```
## what it looked like before left padding
# 1422  ../data/haum_logs_2016-2023/_2016b/2016_11_15-12_2_57.log 2016-12-15 12:12:56  599671 Transform start     076 076.xml   NA 2092.25 2008.00 0.3000000   13.26874254
# 1423 ../data/haum_logs_2016-2023/_2016b/2016_11_15-12_12_57.log 2016-12-15 12:12:57     621 Transform start     076 076.xml   NA 2092.25 2008.00 0.3000000   13.26523465
# 1424 ../data/haum_logs_2016-2023/_2016b/2016_11_15-12_12_57.log 2016-12-15 12:12:57     677  Transform stop     076 076.xml   NA 2092.25 2008.00 0.2997736   13.26239605
# 1425 ../data/haum_logs_2016-2023/_2016b/2016_11_15-12_12_57.log 2016-12-15 12:12:57     774 Transform start     076 076.xml   NA 2092.25 2008.00 0.2999345   13.26239605
# 1426 ../data/haum_logs_2016-2023/_2016b/2016_11_15-12_12_57.log 2016-12-15 12:12:57     850  Transform stop     076 076.xml   NA 2092.25 2008.00 0.2997107   13.26223362
# 1427  ../data/haum_logs_2016-2023/_2016b/2016_11_15-12_2_57.log 2016-12-15 12:12:57  599916  Transform stop     076 076.xml   NA 2092.25 2008.00 0.2997771   13.26523465

## what it looks like now
# 1422 2016_11_15-12_02_57.log 2016-12-15 12:12:56  599671 Transform start     076 076.xml   NA 2092.25 2008.00 0.3000000   13.26874254
# 1423 2016_11_15-12_02_57.log 2016-12-15 12:12:57  599916  Transform stop     076 076.xml   NA 2092.25 2008.00 0.2997771   13.26523465
# 1424 2016_11_15-12_12_57.log 2016-12-15 12:12:57     621 Transform start     076 076.xml   NA 2092.25 2008.00 0.3000000   13.26523465
# 1425 2016_11_15-12_12_57.log 2016-12-15 12:12:57     677  Transform stop     076 076.xml   NA 2092.25 2008.00 0.2997736   13.26239605
# 1426 2016_11_15-12_12_57.log 2016-12-15 12:12:57     774 Transform start     076 076.xml   NA 2092.25 2008.00 0.2999345   13.26239605
# 1427 2016_11_15-12_12_57.log 2016-12-15 12:12:57     850  Transform stop     076 076.xml   NA 2092.25 2008.00 0.2997107   13.26223362
```

## Timestamps repeat

The time stamps in the `date` variable record year, month, day, hour,
minute and seconds. Since one second is not a very short time interval for
a move on a touch display, this is not fine grained enough to bring events
into the correct order, meaning there are events from the same log file
having the same time stamp and even events from different log files having
the same time stamp. The log files get written about every 10 minutes
(which can easily be seen when looking at the file names of the raw log
files). So in order to get events in the correct order, it is necessary to
first order by file ID, within file ID then sort by time stamp `date` and
then within these more coarse grained time stamps sort be `timeMs`. But as
explained above, `timeMs` can only be sorted within one file ID, since they
do not increase consistently over log files, but have a new setoff for each
raw log file.

## x,y-coordinates outside of display range

The display of the Multi-Touch-Table is a 4K-display with 3840 x 2160
pixels. When you plot the start and stop coordinates, the display is
clearly to distinguish. However, a lot of points are outside of the display
range. This can happen, when the art objects are scaled and then moved to
the very edge of the table. Then it will record pixels outside of the
table. These are actually valid data points and I will leave them as is.

```{r}
par(mfrow = c(1, 2))
plot(y.start ~ x.start, dat2)
abline(v = c(0, 3840), h = c(0, 2160), col = "blue", lwd = 2)
plot(y.stop ~ x.stop, dat2)
abline(v = c(0, 3840), h = c(0, 2160), col = "blue", lwd = 2)

aggregate(cbind(x.start, x.stop, y.start, y.stop) ~ 1, dat2, mean)
```

## Pop-ups from glossar cannot be assigned to a specific artwork

All the information, pictures and texts for the topics and pop-ups are
stored in
`/Logfiles/ContentEyevisit/eyevisit_cards_light/<artwork_number>`. Among
other things, each folder contains XML-files with the information about any
technical terms that can be opened from the hypertexts on the topic cards.
Often these information are artwork dependent and then the corresponding
XML-file is in the folder for this artwork. Sometimes, however, more
general terms can be opened. In order to avoid multiple files containing
the same informatione, these were stored in a folder called `glossar` and
get accessed from there. The raw log files only contain the path to this
glossar entry and did not record from which artwork it was accessed. I
tried to assign these glossar entries to the correct artworks. The (very
heuristic) approach was this:

1. Create a lookup table with all XML-file names (possible pop-ups) from
   the glossar folder and what artworks possibly call them. This was stored
   as an `RData` object for easier handling but should maybe be stored in a
   more interoperable format.

2. I went through all possible pop-ups in this lookup table and stored the
   artworks that are associated with it.

3. I created a sub data frame without move events (since they can never be
   associated with a pop-up) and went through every line and looked up if
   an artwork and a topic card had been opened. If this was the case and a
   glossar entry came up before the artwork was closed again, I assigned
   this artwork to this glossar entry.

This is heuristic since it is possible that several topic cards from
different artworks are opened simultaneously and the glossar pop-up could
be opened from either one (it could even be more than two, of course). In
these cases the artwork that was opened closest to the glossar pop-up has
been assigned, but this can never be completely error free.

And this heuristic only assigns a little more than half of the glossar
entries. Since my heuristic only looks for the last artwork that has been
opened and if this artwork is a possible candidate it misses all glossar
pop-ups where another artwork has been opened in between. This is still an
open TODO to write a more elaborate algorithm.

All glossar pop-ups that do not get matched with an artwork are removed
from the data set with a warning.

## Assign a `case` variable based on "time heuristic"

One thing needed in order to work with the data set and use it for machine
learning algorithms like process mining is a variable that tries to
identify a case. A case variable will structure the data frame in a way
that navigation behavior can actually be investigated. However, we do not
know if several people are standing around the table interacting with it or
just one very active person. The simplest way to define a case variable is
to just use a time limit between events. This means that when the table has
not been interacted with for, e.g., 20 seconds than it is assumed that a
person moved on and a new person started interacting with the table. This
is the easiest heuristic and implemented at the moment. Process mining
shows that this simple approach works in a way that the correct process
gets extracted by the algorithm.

In order to investigate user behavior on a more fine grained level, it will
be necessary to come up with a more elaborate approach. A better, still
simple approach could be to use this kind of time limit and additionally
look at the distance between artworks interacted with within one time
window. When artworks are far apart is seems plausible that more than one
person interacted with them. Very short time lapses between events on
different artworks could also be an indicator that more than one person is
interacting with the table.

## Assign a `trace` variable

The `trace` variable is supposed to show one interaction trace with one
artwork. Meaning it starts when an artwork is touched or flipped and stops
when it is closed again. It is easy to assign a trace from flipping a card
over opening (maybe several) topics and pop-ups for this artwork card until
closing this card again. But one would like to assign the same trace to
move events surrounding this interaction. Again, this is not possible in an
algorithmic way but only heuristically. I used the `case` variable in order
to get meaningful units around the artworks.

If within one case only a single trace for a single artwork was opened, I
assigned this trace to the moves associated with this artwork. I (quite
often) happens that within one case one artwork is opened and closed
several times, each time starting a new trace. I then assigned all the
following move events to the trace beforehand. This is, of course,
arbitrary and could also be handled the other way around.

Another possibility is, that an artwork gets moved within one trace without
being flipped. I then assigned a new trace to this move.

This overall worked very well even though it was based on the very
heuristic approach assigning a case when the table has not been touched for
20 seconds. It should be kept in mind that the trace assignments for the
moves will change when case is defined in a different way.

## A `move` event does not record any change

Most of the events in the log files are move events. Additionally, many of
these move events are recorded but they do not indicate any change meaning
the only difference is the time stamp. All other variables indicating moves
like `x.start` and `x.stop`, `rotation.start` and `rotation.stop` etc. do
not show any change. They represent about 2/3 of all move events. These
events are probably short touches of the table without an actual
interaction. They were therefore removed from the data set.

## Events that only close (`date.start` is NA)

It looks like there is some kind of log error for the events that do not
have a start stop. I was able to get rid of most by sorting for `popup` for
the openPopup events, but there are still some left (50 for the small data
set, which corresponds to 0.2 per mill). The following example shows that
artwork "501" gets closed (line 31030) while the pop-up `sommerbau.xml`
is still opened (line 31027). Then artwork "501" gets opened again
(line 31035) and after that the pop-up `sommerbau.xml` is closed (line
31040). This should not be possible and therefore (correctly) two events
are assigned: One where the pop-up was opened and then not closed (which is
common) and another one where the pop-up has no start.

```{r}
dat[31000:31019,]
# Card gets flipped closed before pop-up closes --> log error!
```

I did not check all of these cases (for the complete data set this is
simply not possible by hand) but just excluded all events that do not have
a `date.start` since they are hard to interpret. Often they are log errors
but in some cases they might be resolvable.

```{r}
# remove all events that do not have a `date.start`
dim(dat2[is.na(dat2$date.start), ])
dat2 <- dat2[!is.na(dat2$date.start), ]
```

## Card indices go from 0 to 7 (instead of 0 to 5 as expected)

See `questions_number-of-cards.R` for more details.

I wrote a function that for each artwork extracts the file names of the
possible topic cards and then looks up which topics have actually been
displayed on the back of the card. I added an index giving the ordering in
the index files.

The possible values in the variable `topicNumber` range from 0 to 7,
however, not artwork has more than six different numbers. So I just renamed
those numbers from 1 to the highest number, e.g., $0,1,2,4,5,6$ was changed
to $0\to 1,1\to 2,2\to 3,4\to 4,5\to 5,6\to 6$. Next I used the index to
assign topics and file names to the according pop-ups. This needs to be
cross checked with the programming, but seems the most plausible approach
with my current knowledge.

## Extracting topics from `index.xml` vs. `<artwork_number>.xml`

When I extract the topics from `index.html` I get different topics, than
when I get them from `<artwork>.html`. At first glance, it looks like using
`index.html` actually gives the wrong results.

```{r}
artworks <- unique(dat2$artwork)
path <- "data/haum/ContentEyevisit/eyevisit_cards_light/"
topics <- extract_topics(artworks, rep("index.xml", length(artworks)), path)
topics2 <- extract_topics(artworks, paste0(artworks, ".xml"), path)

topics[!topics$file_name %in% topics2$file_name, ]
topics2[!topics2$file_name %in% topics$file_name, ]
```

For artwork "031", `index.html` only defines 5 cards (the 6th is commented
out), but `topicNumber` for this artwork has 6 different entries. I will
therefore extract the topics from `<artwork>.html`. (This seems also better
compatible with other data sets like 8o8m.)

## New artworks "504" and "505" starting October 2022

When I read in the complete data frame for the first time, all of the
sudden there were 72 instead of 70 artworks. It seems like these two
artworks appear on October 21, 2022.

```{r}
dat0 <- read.table("data/haum/raw_logfiles_2023-09-23_01-31-30.csv",
                   sep = ";", header = TRUE)
dat0$date <- as.POSIXct(dat0$date)
dat0$glossar <- ifelse(dat0$artwork == "glossar", 1, 0)

# Remove irrelevant events
dat <- subset(dat0, !(dat0$event %in% c("Start Application",
                                        "Show Application")))

summary(dat[dat$artwork %in% c("504", "505"), ])
```

The artworks seem to be have updated in general after October 21, 2022.

```{r}
art_after_oct2022 <- sort(unique(dat[dat$date >= "2022-10-21", "artwork"]))
art_before_oct2022 <- sort(unique(dat[dat$date <= "2022-10-21", "artwork"]))
# Removed artworks
art_before_oct2022[!art_before_oct2022 %in% art_after_oct2022]
# Additional artworks
art_after_oct2022[!art_after_oct2022 %in% art_before_oct2022]
```

The following table shows which artworks were presented in which years.

```{r}
xtabs(~ artwork + lubridate::year(date), dat)
```

It strongly suggests that the artworks haven been updated after the Corona
pandemic. I think, the table was also moved to a different location at that
point. (Check with PG to make sure.)

I need to get the XML files for "504" and "505" from PM in order to extract
information on them for the metadata.

# Optimizing resources used by the code

After I started trying out the functions on the complete data set, it
became obvious (not surprisingly `:)`) that this will not work --
especially for the move events. The reshape function cannot take a long
data frame with over 6 Million entries and convert it into a wide data
frame (at least not on my laptop). The code is supposed to work "out of the
box" for researchers, hence it *should* run on a regular (8 core) laptop.
So, I changed the reshaping so that it is done in batches on subsets of the
data for every `fileId` separately. This means that events that span over
two raw log files cannot be closed and will then be removed from the data
set. The functions warns about this, but it is a random process getting rid
of these data and seems therefore not like a systematic problem. Another
reason why this is not bad, is that durations cannot be calculated for
events across log files anyways, because the time stamps do not increase
systematically over log files (see above).

I meant to put the lists back together with `do.call(rbind, some_list)` but
this can also not handle big data sets. I therefore switched to
`dplyr::bind_rows(some_ist)` which is really fast and was developed
especially for this purpose. It means, that I have to depend on the dplyr
package (which I am not a big fan of, since I meant to keep the package
self-contained).

# Reading list

* @Arizmendi2022 [--]
* @Bannert2014 [x]
* @Bousbia2010 [--]
* @Cerezo2020
* @GerjetsSchwan2021 [x]
* @Goldhammer2020
* @Guenther2007
* @HuberBannert2023 [x]
* @Kroehne2018
* @SchwanGerjets2021 [x]
* @vanderAalst2016 [Chap. 2, x]
* @vanderAalst2016 [Chap. 3]
* @vanderAalst2016 [Chap. 5, x]
* @Wang2019

# Open stuff

* Angle from which people approach table in Braunschweig? Consider in
  rotation variable?
* Time limit for `case` variable different for different events? (openTopic
  should be opened the longest)

  $\to$ I think this is not relevant since I am looking at time *between*
  events!

# Stuff AK found interesting

* Pre/post corona
* Identify school classes
* How many persons are present at the table?

# Other potential questions

* "Bursts"
* 1st vs. 2nd half of the day
* Can we identify "types of art"? With clustering or something?
* Possible to estimate how many persons per day? Maybe average of certain
  weekdays? ... ?