Welcome to the Label Maker Demo

Interactive Label Generation for Time Series Data

About

This project is used to evaluate the label maker demo. After successfully installing and using the project please send the resulting database to johannes.windischbauer@tuwien.ac.at and complete the survey.

Installation

Prerequisites

Docker needs to be installed and running on the machine.

Docker

The easiest way to get the application up and running is to start Docker and in the project root run:

docker compose up --build

The application should become available shortly after at localhost:8501

Build Time Information

Duration	Machine	Processor	Memory	OS	incl. download
~13 minutes	MacBook Pro, 13-Inch, 2018	2,3 GHz i5	16GB RAM	macOS Sonoma 14.1.1	yes
5.4 seconds	MacBook Pro, 16-Inch, 2021	Apple M1 Max	64GB RAM	macOS Sonoma 14.0	no

Build From Source on MacOS (M1)

The frontend is using Python 3.11.6 while the backend utilizes 3.9.6. The setup needs to be completed for both the frontend and the backend. To build from source a poetry installation is also needed.

Set poetry version to

poetry env use <python.version | path_to_version>

Spawn poetry shell

poetry shell

Install dependencies

poetry install

Install poetry dot-env-plugin to be able to use .env files.

poetry self add poetry-dotenv-plugin

Run

The preferred way to run the application is by running docker compose up.

If, however, you want to run individual parts of the application on your machine without virtualization you have to install the poetry dependencies in both subprojects. From there on out you can run start the frontend with:

poetry run streamlit run Home.py

and the backend with:

poetry run python rpc-backend.py

It is also possible to run the frontend or backend in a docker and the other part locally. Depending on your machine, it can make sense to run the rpc-backend locally for more labeling performance.

Getting Started

0. Login with the following credentials.

username: user  
password: default

To get the most out of your time using the label maker the proposed usage is as follows.

1. First, take a look at the Sample Labeling Task.

2. Inspect the preexisting labeling results on the Labeling Results page and have a look at the time series in the table.

3. Afterwards go to the Labeling page. Familiarize yourself with the Labeling Properties and metrics. (A description of the properties and metrics can be found further down on this page.)

4. If you have gained a general understand, move to the Ruleset Creator and create a new rule for an existing ruleset. Once you are done with the rule creation, go back to the Labeling page and see if what has changed.

Repeat Steps 3 and 4 as often as necessary and save your progress in between.

5. Save your progress by clicking Persist Labeling Result.

Finally, when you are comfortable, switch to the Use Case task and create at least one ruleset to label the data.

Note: You can at any point in time set the Gold Label for a time series to improve the significance of the metrics provided to you.

Pages

The current state of the demo tries to show off the functionality and should allow for an evaluation of the use case. When running the application via Docker, it is currently not possible to add labeling tasks with other feature matrices and time series for custom evaluation. You can however do this if you run the application locally. To find out how to structure time series and feature matrices look at ./label-generation-data/resources/.

Labeling Page

The labeling page has two tabs. The first tab will show you the selected labeling task. On this page you can gain a quick overview over the labeling task. You can see how many time series and features are available for labeling, how many gold labels are available, as well as how many labeling results already exist. Furthermore, you can see which labels are available for the task at hand. From there on, you can create a ruleset or if there are already labeling results go to view them.

NOTE: There is a known issue where you have to change the ruleset initially to be able to see the stats. This issue is caused by a dependency that allows to add labels conveniently to a new labeling task. After switching labeling tasks once the issue should not arise again.

The second tab offers some additional expanders where you can create new Labeling Tasks, should you run the application locally.

The data url is the path where the feature matrix is located. Either from the project root should you decide to move the data into the project or as a global path. The path where time series are located should point to a directory containing the time series you provide in the feature matrix. Labeling tasks cannot be edited or removed.

Ruleset Page

On this page you can view your ruleset. You can open the expanders of the different rules to view them and their respective labels. There are three types of rule creators, each offering a varying experience for creating rules. Each one offers the possibility to give a rule a name so referring to the rule can be easier. Rules are editable and removable. The rules inside a ruleset are changeable, however, it is not possible to delete a ruleset entirely.

Rule Creators

Every rule creator has different strengths and might be used for a different purpose. All rule creators offer the possibility to show a downscaled sample of how many time series you chose. On the top of the page there is a rule creator expander and drop down, where you can select which creator should be selected. The selection is cached.

Simple Rule Creator

The simple rule creator offers the possibility to explore the data. Features can be selected and the value slider automatically adapts to the data. When changing feature it sits right at the median. There is also the show statistics expander that offers min, max, median, first and third quantile information.

Advanced Rule Creator

The advanced rule creator enables users to use logical connectives for their rules as can be seen in the screenshot. Since the creator is free text based there are two drop downs on the bottom of the page showing which features are available.

Free Form Rule Creator

This is the most advanced rule creator and basically allows a user to write full labeling functions. The features and available labels are available at the bottom of the page. Labels can either be returned by their name or by their index. Note that ABSTAIN has the index -1.

Time Series List and Thumbnail

The time series and thumbnail component have been changed to display in green true positives and in red false positives in regard to the column they are in.

Labeling

The labeling page contains the labeling properties. The labeling properties expander changes the properties of the applied models. Each change reloads the model. Retraining only happens, when the features or labels changed in the mean time. Each surrogate model has different properties.

After setting the properties and scrolling down there are the labeling results.

Labeling Result

Lastly, if there are labeling results available in the database for a respective task, they can be displayed on the labeling results page. The labeling results can be sorted by their scores and offer the possibility to open each one back up in the labeler.

Labeling Metrics and Properties

Properties

The labeling properties offer two sections, one for each model.

Snorkel Labeling

• Snorkel Model: The snorkel model is used for evaluating the labeling functions/rules. There are two models available, the label model and the majority model. More information about the models can be found here.
• Epochs: The number of epochs to train (where each epoch is a single optimization step)
• Seed: A random seed to initialize the random number generator with
• Tie Break Policy: What the model does in case of a tie. Either abstain, random or true-random (not encouraged)
• Only use selected labeling functions: If the ruleset contains many rules, there is the option to exclude some.

Surrogate Model

• Surrogate Model: Three available surrogate models: A scikit-learn Logistic Regression, a pytorch-lightning driven probabilistic logistic regression and the WeaSEL model using a MLPNet as end-model.
• Restart Surrogate Model Training: In case the input parameters do not change the model will not be retrained. If you want to retrain for any reason, click this button.
• Use Gold Labels (where available): This replaces the labels from the Label Model with the available Gold Labels for training the Surrogate Model.
• Use only the same features as the Snorkel Model: The feature matrices contain many features. Not all of them are representative for the task you want to achieve, therefore, the pre-selection is true. If more features are needed that are not already included via the labeling functions they can be added in the additional features multi-select below.
• C and Solver: Are scikit-learn parameters for the logistic regression. More information can be found here
• Number of Stratified K-Folds: Sets the K, the amount of folds, for an evaluation metric. Has no real influence on training and solely allows to gauge the generalization capabilities.

Metrics

The labeling Results dataframe is the result of the labeling step with the applied labeling properties.

• mkey: A unique key for every metric
• Time Series: A 20x downscaled line chart of the time series.
• Gold Label: The selected gold label for a time series. Can be changed here.
• label model: The label that results from applying the Snorkel Model used for training the surrogate model.
• label model propba: The probability with which the label model label applies.
• surrogate model: The final label after training and predicting on the surrogate model.
• go to: allows a user to inspect a time series closer by advancing to a time series analysis page.

The Labeling Function Analysis includes four sections.

• snorkel lfa: The snorkel labeling function analysis offers insight about each rule. In-depth information here
• Label Model compared to Gold Label: This section calculates the well known metrics accuracy, precision, and recall compared to the available gold labels. This indicates how well the labeling functions fit the task.
• Features used: This is the set of features extracted from the labeling rules that was used for training.
• Labeling Rules used: Shows which labeling rules have been used in json format.

The Surrogate Model Analysis has five sections and focuses on the performance of the Surrogate Model.

• Number of Stratified K-Folds: Shows how well the Surrogate Model performs in regard to the training labels provided by the Label Model. The scores are averaged on the number of folds.
• Gold Label Scores: This shows the same metrics as the previous section compared to the Gold Labels. This is similar to the Label Model compared to Gold Label, but for the Surrogate Model. The Surrogate Model has then been trained on the whole training data and not on a specific fold.
• Label Model Scores: The Label Model Scores show the scores compared to the training label data.
• Features used: Which features have been used for the training. Can be more than in the Labeling Funcion Analysis.
• Confusion Matrix: Shows the confusion matrix compared to the training labels.

Survey

If you have finished the Seasonal or Constant use case please send the database file located at data/label_maker.sqlite to johannes.windischbauer@tuwien.ac.at and fill out the survey. It should take no longer than 10 minutes.

Thank you