Filter by

• wildcards
• slider for k-most-relevant features #78
• manual selecting / hiding #81

jump to feature on map

Arvind wondered why we switched from bubbles to dots because it is hard to hit the tiny dots. Any reasoning for this?

We could add a small question mark button besides the title that shows some description about the information we are dispalying and why

As Bachelor Project Student I want to see want to track the engineers' behavior with our prototype in order to analyze and improve interactions and UX.

In the backend we will have an endpoint POST /api/users/events with the payload {"type": "an_event_name", "options": {...}}. The options dict and the event type are dynamically definied by the Client.

Highlight interesting histogram buckets depending on the relevancies of the slices in them.
This should give the user an impression on where to look first.

@KDD-OpenSource/ba-project-2017 Render client side or server side?

Server side we could just run sth like this http://scikit-learn.org/stable/modules/density.html

I added the following endpoints:

• Send {"username": <USERNAME>, "password": <PASSWORD>} to POST /api/auth/login in order to obtain a token: {"token": <TOKEN>}
• Use the token from above to send authenticated requests by adding the Authorization header with the following value Token <TOKEN>
• Send DELETE /api/auth/logout to logout and delete token from DB
• Each token is associated with a user. It also never expires! Save it somewhere in local storage

Do you think we need a register feature? We can create users pretty easily in the admin panel.

Django now has built-in support for websocket and model changes which I want to make use of. The websocket url changed to /bindings. The payload now looks like this.

{  
   "stream":"rar_result",
   "payload":{  
      "action":"update",
      "pk":"69b0bab5-8f32-4ff5-a6ae-1cd278c75d93",
      "data":{  
         "relevancy":0.2,
         "rank":1,
         "feature":"502cd3b2-b426-4d68-8f60-58d7f93959f9"
      },
      "model":"features.rarresult"
   }
}

Also make sure, that you check for the stream value of rar_result before parsing the payload.

Now, we don't even need to re-pull the results. Note that the model change signals are not triggered, when objects get created from the shell. For testing I added a admin page for rar_results

https://docs.angularjs.org/guide/unit-testing

Write grunt and npm tasks to

• Remove sails
• compile (less, coffeescript, pug / jade)
• serve locally
• hot recompile

because those are not of interest (e.g. not causal)

What changed in the backend

• Basic user support and auth built in but not enabled
• New Dataset object
• New Session object which is unique per user, dataset (fixed after creation) and target feature
• Results from Rar in special results object and not feature itself anymore
• From now on resources need to be fetched by their ID and not their name (e.g. for feature)
• PUT /api/datasets/upload: to upload and create new datasets, also starts tasks for feature initalization. 'file': as header
• GET /api/datasets: Fetch all existing datasets
• GET /api/datasets/<dataset_id>/features Fetch all features from dataset
• GET /api/features/<feature_id>/samples Fetch samples of feature
• GET /api/features/<feature_id>/histogram Fetch histogram buckets of feature
• GET /api/sessions: Get all running session
• POST /api/sessions: Add new session with dataset
• GET /api/sessions/<session_id>: Get specific session including target
• PUT /api/sessions/<session_id>/target: Change target and start rar by sending {"target": <feature_id>}, caches existing rar calculations per target and dataset
• GET /api/sessions/<session_id>/features/<feature_id>/slices: Get slices of a feature with respect to session (dataset, target)
• GET /api/sessions/<session_id>/rar_results: Get rar_results for each feature with respect to session (dataset, target)

I propose to split up the work somehow.

Keep this constraint when the target changes!

Instead of filtering our samples in the frontend, we should get a new sample from the backend after selecting a feature range. That way, we can avoid "not enough data" when plotting the graph.

• GET /api/tasks to get a list of currently running tasks:

{
  "scheduled": {
    "celery@616cf6a06dc3": []
  },
  "active": {
    "celery@616cf6a06dc3": [
      {
        "id": "0758841f-c92b-49bf-a39e-19add218ed09",
        "name": "features.tasks.calculate_rar",
        "args": "()",
        "kwargs": "{'target_id': UUID('a33a0d38-3f0b-44ea-8083-5ccda9df7480')}",
        "type": "features.tasks.calculate_rar",
        "hostname": "celery@616cf6a06dc3",
        "time_start": 60600.905955012,
        "acknowledged": true,
        "delivery_info": {
          "exchange": "",
          "routing_key": "celery",
          "priority": 0,
          "redelivered": false
        },
        "worker_pid": 20
      }
    ]
  },
  "reserved": {
    "celery@616cf6a06dc3": []
  }
}

• DELETE /api/tasks/<ID> to terminate a task

Consider showing a better name than features.tasks.calculate_rar to the user. You can also think about showing a progress bar after initialization. Don't query too often as the request takes more than 3s.

Additionally to the "Set target" option in the hamburger menu I would add a button "Set as target" to each individual feature overview

I want to confirm a point.
Imagine I have feature A, B and a target class Y.
If A is redundant to B and both are relevant to Y, then A and B will be placed on two different hemispheres of the circular structure right?

Abstractly my question is: If I divide the circle vertically in to two hemispheres. On what basis do you decide, where to put the features (in left hand side or right hand side of the hemisphere)?

• Neue Websocket streams relevancy_result und redundancy_result
• is_categorical Attribut für Feature
• Neue/andere Endpoints

 +    url(r'targets/(?P<target_id>[a-zA-Z0-9-]+)/features/(?P<feature_id>[a-zA-Z0-9-]+)/slices$',
          FeatureSlicesView.as_view(),
 -        name='session-feature-slices'),
 -    url(r'sessions/(?P<session_id>[a-zA-Z0-9-]+)/rar_results$',
 -        FeatureRarResultsView.as_view(),
 -        name='session-feature-rar_results'),
 +        name='target-feature-slices'),
 +    url(r'targets/(?P<target_id>[a-zA-Z0-9-]+)/relevancy_results$',
 +        FeatureRelevancyResultsView.as_view(),
 +        name='target-feature-relevancy_results'),
 +    url(r'datasets/(?P<dataset_id>[a-zA-Z0-9-]+)/redundancy_results$',
 +        DatasetRedundancyResults.as_view(),
 +        name='feature-redundancy_results'),
 +     url('targets/(?P<target_id>[a-zA-Z0-9-]+)/slices$', FilteredSlicesView.as_view(),
 +        name='target-filtered-slices'),```

Use checkboxes instead of sliders in this screen.

http://plottablejs.org/ from Palantir, useful?

The scatterplot should have the currently sliced out points highlighted (linking & brushing)

Fix or create popup that says pls use chrome

Search for feature textbox at the top left corner not responsive. When I type a feature name in the textbox and it gives me 2 options in the drop down.

1. Set the feature as target
2. locate the feature.

I click on locate the feature, but it is not responding.
Its just doing nothing. I thought it should ideally add it into the list of selected features.

Send token with initial websocket request to be authenticated