Creating an empty plugin

To create a new plugin, click the Custom menu at the top and select Plugins at the bottom. First give your plugin a name such as Clustering. The description and helper documentation are optional and can be supplied later.

Then choose the input data. For intelligence plugins you always choose Intelligence model.

From Intelligence goal choose None if your plugin doesn't work with some conversion goal, for example an unsupervised clustering plugin. In all other cases, such as a prediction plugin, choose Binary.

From Initial dataset moment pick the relative time for which you want a snapshot of the dataset of users. This is relevant for features that count the number of times someone did an action, like song played. For moment Initial state of user, the number of songs played will always be 0, because the user didn't have to yet to make any actions — we gave them 0 seconds of time since their existence. Latest state of user means that we count all actions up to the current time.

Moment where 95% of users didn't convert yet' means that we first calculate for all users the time in seconds it took them to convert (since their creation), then take the 5th percentile of that. Then we use the resulting 5th percentile as number of seconds since user creation for the dataset, to measure only actions done up to that number of seconds. This means that if we only want to analyze users that didn't convert yet, we have thus a dataset of 95% of the users, because 5% will be discarded as they've already converted and are not useful for our analysis anymore. While 95% of users is a good rule of thumb we provide for convenience, you can request more dataset snapshots from any time in seconds since creation, or any other % of users, once the initial processing round is done. More on that in the Initial results section.

Next we can choose whether Plugin only works with unbiased user properties. By default this is selected, and means that any user properties not defined at user creation are removed from the analysis. Say we have a user property renewed of boolean type, indicating whether the user has renewed another year of their paid plan. When we do an analysis to predict what user properties lead to payment conversion, this property will surely have a very strong correlation with the payment conversion goal. To prevent such biased properties from being used, but keep only the ones that were set already at user creation time (such as country and device), check [✅] Plugin only works with unbiased user properties. To mark any user property unbiased, visit your project settings page and click Set-Up Data, then under the Data Management tab check the box for any unbiased properties.

Note that first visit user properties such as country or device type are always available, independent of the dataset snapshot time, because those are part of user creation and thus available at 0 seconds already.

The code will be left untouched for now; you will first develop the plugin locally using your own toolchain. Currently only Python 3.7 code is supported, but more languages will be available in the near future.

The last step is to [✅] Run plugin in test-mode and then Add Plugin.

Now that your plugin is added to the template, save your template. Click Save Changes from the top menu, and give your template a title. In case you want to make your template and any plugins within it available publically to others, check [✅] Make publicly available as third-party template. Finally click Save to save your template.