the-strategy-unit / experiencesdashboard Goto Github PK

https://github.com/CDU-data-science-team/experiencesdashboard/blob/9e1a55b1410abba6e98bb2445e7f9cc60deabe7a/R/mod_sentiment.R#L217

the example in this discussion doesnt seem to work in golem see rstudio/shiny#650

probably not a good idea to hardcode all of these input boxes, check with @ChrisBeeley

https://github.com/CDU-data-science-team/experiencesdashboard/blob/9e1a55b1410abba6e98bb2445e7f9cc60deabe7a/R/mod_sentiment.R#L31

https://github.com/CDU-data-science-team/experiencesdashboard/blob/9e1a55b1410abba6e98bb2445e7f9cc60deabe7a/R/mod_sentiment.R#L50

https://github.com/CDU-data-science-team/experiencesdashboard/blob/9e1a55b1410abba6e98bb2445e7f9cc60deabe7a/R/mod_sentiment.R#L95

using the same settings for filtering I get different results

for example one module for each plot / table

Consider either removing the FFT and demographics tab from the demo, or adding in fake data automatically to make them work

this is primarily for visualisations, but could also be relevant for text

Offer drilldown by location_2 and location_3

Improve text searching- allow multiple selections, excluding text, and maybe a semantic search (this could even use document vectors if we want to get really clever)

not sure why this is happening, this is the error message

Warning: Error in if: argument is of length zero
  [No stack trace available]

Real demographic data from trust C is not working well. Clean and fix

I'd like to be able to generate customisable reports in the app:

1. first make some selections (e.g. date range, division, ...)
2. then apply all these selections to differents parts of the shiny app that the user selects e.g.,

• user1 selects these elements from the shiny app

  • generate table from tab1 (e.g. patient experiences)
  • generate plot from tab2 (e.g. staff experiences)
  • generate text summarising tab3 (e.g. complaints)

• user2 selects these elements from the shiny app

  • generate plot from tab1 (e.g. patient experiences)
  • generate table from tab2 (e.g. staff experiences)
  • generate text summarising tab4 (e.g. compliments)

do you think this is a good idea @ChrisBeeley ? and do you think it's possible?

It would be useful if users could download comments that are likely to be insufficiently described by one code so they can be manually reviewed. Examples of comments that would be included are:

• Long comments
• Comments where the document vector representations of the sentences within them are far apart
• (if we adopted a probabilistic ML model) comments where there isn't a clear winner for which category they belong to

Produce an integrated report that summarises numbers of different demographic groups in the survey as compared with the number in the service and the population well as different scores and comments for each demographic group

When I run the code from README (or even try to build on my laptop) I get the following error:

devtools::install_github("CDU-data-science-team/experiencesdashboard")

Error

Downloading GitHub repo CDU-data-science-team/experiencesdashboard@HEAD
Error: Failed to install 'experiencesdashboard' from GitHub:
  Missing commas separating Remotes: 'CDU-data-science-team/nottshcMethods
    CDU-data-science-team/experienceAnalysis@development'

Local build error:

ERROR: dependencies 'dashboardthemes', 'waiter', 'shinyjqui', 'timevis', 'UpSetR', 'tibbletime', 'tidytext', 'experienceAnalysis', 'reticulate', 'Rdpack' are not available for package 'experiencesdashboard'

this wording is confusing because we use it in the positive (what was good) and negative (what could be done better) context.

something like "weighting" or "importance" might be better, let's discuss @ChrisBeeley @andreassot10

As can be seen in the following places:

754870a

dc331b6

2819195

I can't get the Rmd to be found when the package is installed from GitHub. I've tried moving it to inst/ and to inst/app/www and have not been able to resolve.

I would love some help with it, I feel like I've tried everything

It would be useful if users could interact with a comment that is badly tagged and send that information back to us. Probably the control would just open an email with the text in, nothing more complex than that

Summarise the proportion of each type of feedback, and allow users to click categories to see examples of text

At present the sentiment/ weighting tab allows selection of themes which are not present in the data, and doesn't always land on a combination of themes which are present in the data. Themes which don't exist in the current selection should be removed so users can't click on things that don't do anything

#7 references reports but is quite a complicated implementation of them in the 0.3 release. We need some simple reports in the 0.2 release so the partner trusts can use and review example reports.

They could just consist of:

• an SPC for the FFT question (see #17 )
• A table summarising the % of each theme
• A table summarising the +ve -ve level of criticality/ sentiment
• Maybe a table that cross tabs theme and criticality/ sentiment

It will kick off a conversation even if it's not quite what they want

There was a downloadable version, but it doesn't work now because the app works with MySQL rather than data/.

Add a downloadable version

I think we're missing a trick by not prototyping with {shinipsum} https://cran.r-project.org/web/packages/shinipsum/index.html

I'm going to rebuild the whole thing with a data module for #5 and throw up some wireframes with shinipsum. This will help us to at least have placeholder stuff in when we talk to our partners, even if the output isn't built. Then we can talk about the structure and flesh out the components at the same time.

Start simple with search by keyword

I've loaded the package and dependencies and when I run the app it appears but then shuts down with an error:

> shiny::runApp()
Loading required package: shiny
Warning in loadSupport(appDir, renv = sharedEnv, globalrenv = NULL) :
  Loading R/ subdirectory for Shiny application, but this directory appears to contain an R package. Sourcing files in R/ may cause unexpected behavior.
i Loading experiencesdashboard

Listening on http://127.0.0.1:4255
Warning: Error in : nanodbc/nanodbc.cpp:1021: IM002: [Microsoft][ODBC Driver Manager] Data source name not found and no default driver specified 
  62: <Anonymous>
Error : nanodbc/nanodbc.cpp:1021: IM002: [Microsoft][ODBC Driver Manager] Data source name not found and no default driver specified 

I've tried reloading the package when on the VPN as that sometimes helps but I still get the same error.

SPC XMR type charts are required- the default should be monthly, but there may need to be extra options that can be used when there are fewer data points available.

Everyone loves word clouds 😄

(well, most people do 😉 )

Let's include word clouds somewhere too

The experience data for nottshc is now available from the {nottshc} package so it would be good to use that to load the data from.

In theory the data could be pulled live off the server rather than relying on .rda files as at present but we'd have to see what this did to the speed of the app (it may speed it up since otherwise the full version will download 12 years of data)

See CDU-data-science-team/pxtextminingdashboard#36

I was thinking an obvious improvement to the upset plot would be interactivity. Clicking on it would either bring up the relevant comments or flick to the comment tab and pre populate the controls to show those sentiments.

Interactivity seems difficult/ impossible (?) with UpsetR and ggupset but looks simple with {upsetjs} https://cran.r-project.org/web/packages/upsetjs/

I don't really understand upset plots but if you can plot it with upsetjs I can handle the interactivity?

especially in visualisations, e.g. create a factor that is ordered positive first, then all negative

In live use users are likely going to want to add new data, which they will need to do with a spreadsheet in a particular format.

The button will upload the data, run it through the classification algorithm via reticulate, and then write the data back to the database. This issue will add the button and allow nonsense categories to be written to the database, a linked issue will add the reticulate pipeline

Many filter boxes automatically populate with random data, this was done for dev purposes and is now unhelpful. All filters should start empty and default to returning all data

Lots of trusts seem to have SPC charts for the ticky box FFT question (although ours doesn't). We should add it as an option.

For the purposes of the repo we can make fake FFT data because we don't need to publish actual FFT data on here

It would be useful to have a dummy version of the application in the cloud so anybody could put their data in and see the results come back

create categories, for example if user looks up word "meal" the app could also look for the terms "food" and "drink"

Upload functionality only currently works for single comments. Extend it to allow multiple comments to be uploaded and tagged simultaneously

Users should be able to find similar comments to the one they are reading by clicking. This could be done using something like the soft cosine measure https://radimrehurek.com/gensim/auto_examples/tutorials/run_scm.html

this is needed to build a function that allows to compare the feedback/sentiment between two (or more) teams

The dates in the sentiment analysis are faked because the date ranges of the two datasets don't match up. The sentiment analysis needs to be moved into the data module in #5 or perhaps it will come from the DB itself as in The-Strategy-Unit/pxtextmining#9

Report number of surveys from different demographic groups and compare with the population in the service and the community population

1 = XXX
2 = XXX
3 = XXX

Changing the range of the criticality variable has broken the nice little coloured pills on the reactable

I did try to fix but I don't really understand the code, can you show me please?

currently I visualise the density (which is affected by the number of responses during a specific time period). @ChrisBeeley suggested to visualise proportions instead.

I haven't thought about it in detail but I think we need a separate data module that does all this stuff- looks at the data in the application and figures out what everything is, what the date column is, where the location columns are, what they're called and how many there are, all that stuff. The data module can feed the other modules data as well as meta data (number of columns, number of categories...)

Originally posted by @ChrisBeeley in #2 (comment)