Designing non-linear navigation for machine learning and topic modeling experiences

Designing non-linear navigation for machine learning and topic modeling experiencesTyraleBlockedUnblockFollowFollowingFeb 5BaselineBreadcrumbs are great.

They are a fantastic tool that will show the relationship through parent/child elements.

Commonly seen in e-commerce when looking through categorical structured information.

I’m looking at pot on target.


I can see it’s parent categories.

Commonly, they are seen in e-commerce when looking through categorical structured information.

I’m looking at pot on target.


I can see its parent categories: kitchen & dining, cookware & bakeware, pots & pans.

When the intent of the user is clear, breadcrumbs provide a valuable tool for displaying the path of the user’s experience.

As a user, I’ve navigated through this selection process.

When and if I return to this item, I would expect to see the same breadcrumb path.

Breadcrumbs are great for structured data.

Root IssueWhen your data is unstructured and has no clear relational connection from one object to another, how can (who? what?) support the navigational path and user experience to reduce the cognitive load of where the user is located?Use case/Problem StatementA user is investigating human communication that has been flagged with Machine Learning and sorted by topic modeling within a department of an organization looking for inappropriate behavior.

As the user is navigating around, there is no clear relationship between the communication objects or groups.

If a user discovers a piece of communication that is important and then wants to save/share/flag this for later user, what is the object’s path?The user’s intent is not the same as the navigational path.

The object has no unilateral parentThe object can be discovered through an unlimited number of pathways.

ThoughtsHow do we visually represent the user’s experience in the navigation?How does the browser back behave vs the user’s expectations/needs?Understanding the problemWhen a user is within an unstructured data environment that was processed with topic modeling, the relationship between the items feels random.

Thanks to Joe Porter for the example.

You could generate data on high school senior book reports on Napoleon Bonaparte.

When you generate the topics, they could be clusters of any type of category, like row B in the figure.

The concepts in row A were also available, based on existing knowledge.

We could have the following possible topics, and their associated topic paths:Napoleon born in Corsica [A.

1 -> B.

1 -> C.

1]Napoleon emperor of France [A.

1 -> B.

2 -> C.

1, A2 -> B.

4 -> C.

1]Napoleon death from cancer [A.

3 -> B.

5 -> C.

1]French health during Napoleon’s reign [A.

3 -> B.

2 -> C.

1]A user might navigate to the concept Napoleon along any of these paths, each of which expresses a different intent.

Different users would likely use different paths to get to the same concept.

A user might also have arrived at Napoleon directly (e.


, by searching), and then have navigated away along any of the available paths.

Framing questionsIf a user as discovered C thru [A.

3 -> B.

5 -> C.

1] and wants to then share this object to a colleague, what is the proper interaction mechanic and display when it is presented to the new colleague?Then you can layer on the user’s intent.

Is the intent that C is important on its own as an object, or only is it only important in relation to how it was discovered within the topic path of [A.

3 -> B.

5 -> C.

1] (assuming single intent)?Is Napoleon important to this user because he is a notable person[C.

1] within France[B.

2] which is a country[A.

1]?.Or is Napoleon important because he is a notable [C.

1] within a General[B.

3] that is a Person[A.


3 -> B.

5 -> C.

1] vs [A.

1 > B.

1 > C.

1]Analyzing the problem further, is C relevant to User 1 as a person in a location, but when shared to User 2 he is important as a person (assuming multiple intent)?Default Browser Back BehaviorKeeping the browser back button behavior intact is crucial.

The browser back will need to behave as intended and how users would expect.

We do not want to work against the familiarity of established patterns.

Take me back to the last thing.

Design Exploration — Routine task vs InvestigationNavigating through data and content are not always linear.

Our design solution needs to support the concept of both navigation and multithreaded workflows.

The user’s intent can be a combination of following interesting data or performing a routine task.

In my specific application, I can group similar topics into intents that I termed — tiers.

Tier 1 (High level view of watched collections — Dashboard)Tier 2 (Collection of similar topics defined by search parameters — Queue)Tier 3 (The display of the data object — Message)Re: Routine task — To support the routine task, we can rely more on search and lists to get the user to familiar topics that need to be addressed.

We can provide a common, reusable place for the user to navigate to their destination.

Sample TDD covering the user cases and scenarios for a dynamic navigation.

Re: Following the data — To support this flow we will drop tier icons, or markers, in the navigation to allow the user to return to the last parent tier they were visiting.

This is not a full historical breadcrumb, but more of a path back to the “path” parent.

When a user’s intent is investigation, the cognitive path becomes more relevant than the “browser history”.

The design should support this with a dynamic and reactive icon-based navigation.

The icons represent the previous collection, group, and/or topic that is related to the current view.

The user’s cognitive path is visualized through the large pivot points in icons.

Icons are used in place of traditional text to reduce the variability of word or phrase length.

If the dataset has any type of collection or topic utilization, our research has shown this reinforces the user’s cognitive history more than the spelled out topic.

Design TestingWe are currently testing this implementation with some of our Early Adopters and Subject Matter Experts.

Our hope is the data and user’s experience will validate the dynamic nature of their investigation.

In Collaboration; A brilliant Data Scientist, Joe Porter.

Who generated the example used, and continues to educate me on data science.

With nearly everything I work on, Josh (Josh Goodwin) is major collaborator.

Much of the thought and time developing these posts and projects were only possible with his contributions.


. More details

Leave a Reply