We're excited to say that we've added a new panel to your projects! Find Models after you've built a prediction model using the Models | Train & Predict analysis type.
Models is designed to help you understand more about the creation and performance of predictions models you build. You'll find three tabs - General - Training - Result - each containing distinct information about your model.
This information helps you unpick how your model was built, its strengths and weaknesses as well as how it performed in specific areas.
As the statistician, George E. P. Box wrote, "All models are wrong, but some models are useful." What he meant by that is that all models are simplifications of the universe, as they must necessarily be."
Nate Silver, The Signal and the Noise
Examining the mechanics behind a model is a crucial aspect of ensuring that it's application is appropriate. We've included references to the technology used to develop the model, descriptions of its primary use cases as well as notes from our team on its performance.
Evaluating a prediction model helps us to understand how good the model is. Using accuracy scores and other performance metrics, we get a sense of whether the model was able to make correct or inaccurate predictions. Not only this, but through evaluation of a model, we can understand how to improve it by changing its factors or parameters.
Prediction can be pretty complex at the best of times. We've written a few articles to help you get your head around the concepts key to Models.
Start here. This article walks you around the new Models panel - pointing out the different features and showing you how to use them.
Learn why and how to build models. Here, we explicate the process of building a prediction model in Graphext - considering the reasons for doing so alongside some of the key concepts involved.
Use as a dictionary to your Models panel. As well as explaining the usefulness of evaluating models, this article offers explanations for all of the technical terms used inside your project's Models panel.
How can I start using it?
Throughout May, we've been pouring our efforts into fixing bugs and making improvements in Graphext's UX. We've redesigned the project setup wizard, cleaning up the types of analysis you can conduct as well as improving key flows. We've also added a new enrichment option that groups similar spellings in text or categorical variables.
We've made substantial improvements to our old Network of Columns analysis type. This flow lets you study the links between variables in your dataset.
Choose Models → Cluster Variables to build a project that maps the relationships between variables in your data.
Clustering your variables can be a useful way to understand which factors to feed into a model or to simply grasp which variables are strongly linked to one another.
When setting up a Cluster Variables project, choosing a target variable means that your project will focus on mapping other variables relationships to that target variable.
- Choose a dataset with at least two variables.
- Select Models → Cluster Variables as your analysis type.
- Pick which variables you want to cluster in your project.
- Refine your configuration using the questions in the setup wizard.
- Execute your project and study the relationships between your variables.
Aimed at improving the way you conduct Text Analysis in Graphext, our latest data enrichment groups words with similar spellings. Simply put - the idea is to stop Graphext and Graphex from being considered as two separate entities.
Whether it be typos, misplaced punctuation or a missing letter or two, unintended variation in data is a common - and annoying - occurrence in text analysis. Motivated to overcome this common shortcoming, our team of data scientists and engineers built this algorithm to merge words with similar spellings and made it instantly deployable in Graphext using any type of analysis.
Chose Group similar spellings from the list of enrichment options in your data enrichment tab to start grouping similar text or categorical values. Then, set a threshold to configure the strength of the merges taking place.
- Start building a project using a dataset with a text column.
- Choose an analysis type and open the data enrichment tab.
- Select Group similar spellings from the list of enrichment options.
- Set a threshold to control the strength of your word joinings.
- Continue building your project.
- Open your project and check out the new merged variable.
Redesign of the project setup wizard. Without removing any of our capabilities, we've tidied up the way that flows are presented. We've removed Employees and Survey analysis types and renamed Google Analytics to Marketing Attribution. You can build the same project using the Models analysis type. We've done this to make it simpler to find the right kind of analysis for your project.
- Fixed a bug preventing users from segmenting data using a direct selection of nodes in the Graph.
- Fixed a bug stopping users on some Mac OS from extracting CSV files downloaded from Graphext.
- Fixed an issue causing some minor Graph UI features to overlap on Safari browsers.
- Disabled a users ability to create insights inside of projects embedded on external websites.
- Fixed a bug stopping users from changing the color of a segmentation whilst - at the same time - renaming the segmentation.
Attempting to discover the most influential features of a loan application when considering risk, our team built a model using the features of a loan application to predict whether an applicant would have a good or bad risk rating.
In this guide, María and Paul walk you through the process of building a prediction model that analyzes a dataset of 5110 healthcare patients. The model we help you to build will use factors detailing the lifestyle and existing health conditions of a person in order to predict the likelihood of that person suffering a stroke.
We've been improving the flow of key Graphext features to make them more instinctive to work with. You can now save insights to your recipe so that you don't lose them when you recreate a project. It's also easier to save and edit new segmentations. We're also getting ready to introduce you to some substantial - and colorful - new features next month!
Insights are key. They help store your discoveries and build data-driven narratives. You don't want to lose them if you recreate a project using a different flow. We've added the ability to save insights to your recipe. Choose key aspects of your analysis, toggle the save switch and move your analysis forward without losing your findings.
Insights that you save will appear in the Insights panel of your recreated project. You can change the configuration of your recreated project as much as you like - this won't affect the insights that you save.
- Start by saving some insights in a project you have built.
- Then, click the green recipe icon at the bottom of the insights that you want to save.
- Toggle the switch to save the insight to your recipe.
- Recreate the project using the settings menu located on the top left of your screen.
- Edit your projects configuration until you are happy with the changes. Execute it.
- Head over to the new project's insights panel to inspect your saved insights.
Creating manual or automatic segmentations is a powerful way to discover sub-communities in your data. We've made it easier to customize the properties of segmentations by improving the logic with which you edit a segmentation. We've also added a button to help you cancel the changes that you make and made it simpler to undo steps like renaming or coloring segments.
Segmentations act like a new variable in your dataset, dividing your values along lines that let you see your data from different perspectives. Segmentations are communities in your data and it is important to control how they appear.
Changes we've made to the flow of editing a segmentation are designed to give you greater freedom in renaming, coloring or removing segmentations. Make use of the undo and cancel buttons to control your changes precisely and quickly.
- Create a segmentation in one of your Graphext projects.
- Click on the Edit Segmentation icon from inside the segmentations sidebar card.
- Use the icons to rename, color and remove segments.
- Undo and cancel any changes you make using the icons next to the Save button.
- Fixed a bug preventing sub-cluster region labels from appearing after sub-clustering segments.
- Fixed an issue with the product basket analysis flow.
- Fixed an issue with the network of columns flow.
- Fixed a bug in the keywords flow when using datasets containing URL variables.
- Improved the display of categorical values in your sidebar variable charts. Now - no matter the number of values - categorical variables are always presented in bar charts, not lists.
Andy and María met with Jake to talk about a dataset he's building about himself. From skating to people he sees to whether he eats meat or not - Jake's data diary offers a unique and deeply personal insight into his life. But what makes the difference between good and bad days?
Businesses fighting to understand feedback from hundreds or thousands of customer reviews can analyze what people are saying using NLP techniques but it takes time and resources to do so. This guide is intended to walk you through the process of analyzing customer reviews with Graphext. We will analyze a dataset of 42,656 reviews about 3 Disneyland branches using the Text → Topics flow.
This month, we've been working on making Graphext easier and more intuitive to use. We've added features making it easier to spot relationships between variables using compare charts and inspect data on your Graph quickly by controlling size mapping with greater precision.
We've also added a new type of analysis making it possible to analyze the relationships between recurring items in your data. On top of this, we've extended the list of language support options we have so that you can analyze text written in Turkish and Arabic.
It just got much easier to start analyzing the variables highlighting relationships in your data when you build a prediction model in Graphext.
To make it faster for you to spot key variables, we've re-configured the way that the Compare panel generates charts explaining differences or similarities between values in your dataset.
Now, when you generate compare charts after building a model using the train and predict analysis type, Graphext will automatically display only important variable charts, hiding the variable charts that aren't as relevant to your analysis. You can change this by switching between the new categories of charts that we've created.
Use the dropdown menus at the top of your Compare panel to start toggling between these variable collections.
All - All variables
Target - Modelled variable
Factors - Variables used to create the model
Other Variables - Variables not used as factors or target
Important - Target(s) and Factors
Internal Variables - Variables created by Graphext
None - Select a variable individually
- Start by building a project using the train and predict analysis type.
- Then, from the Compare panel of your project, choose some variables to compare.
- Graphext will automatically display the Important collection of variable charts.
- Toggle between collections using the dropdown menu at the top of your Compare panel.
Node sizes are a great way of exploring quantitative values in your Graph at a glance. You can now control the range of sizes that nodes are given using a sliding scale.
Select the node size icon from the icon collection at the top of your project's Graph to start customizing the range of node sizes presented in your project. You can control the top and bottom of the range using the slider.
You can also control your node sizes from your project settings. If you want to save your node size configuration, you can do this using the node size slider inside of the project settings window.
- Open your project and navigate to the Graph.
- Select the node size icon from the icon list at the top of your Graph.
- Click the three dots from the top right of the variable card.
- Move the range presented in the slider to change the size of your nodes.
We've added a new type of analysis called Co-occurrence. You can find it within the Models section of the project setup wizard.
Co-occurrence analysis lets you find relationships between recurring items in your dataset. It helps you identify which items are most associated with other items.
Whilst it's already possible to conduct co-occurrence analysis with Text and Product Basket analysis types in Graphext, we built this flow to work with any kind of data. Use Models → Co-occurrence analysis to discover the associations between a range of entities which might be people, products or places.
- Upload a dataset containing recurring items.
- Start building your project using the setup wizard.
- Choose Models as your type of analysis.
- Choose Co-occurrence as your sub-analysis type.
- Tell Graphext how you want to aggregate your data and which column contains your items.
- Execute the project and start discovering how items are related.
Our team of engineers and data scientists have been busy improving Graphext's Natural Language Processing capacity so that it is possible to analyze content written in more languages using our text analysis flows.
You can now analyze text written in Arabic and Turkish on top of the existing language support options we already have built-in. We've also made it much easier to extend the list of languages we support!
Read more about language support at Graphext here.
- Start with a dataset containing at least one text field and choose Text or Social Media as your analysis type using the project setup wizard.
- Inside of the Data Extraction tab, choose how you would like to set the language of your text.
- You can set languages manually or by inferring it directly from the text itself.
- That's it. Execute your project and delve into your analysis.
- Fixed a bug causing the building of projects to fail after a user decided to delete data points and recreate the project without these points.
- Solved issues surrounding an inability to save new manual segmentations.
- We've made it easier to build projects using datasets that match our built-in analysis types. When you build a project matching an analysis type, Graphext will make assumptions on the way you want to set up your project. You can still edit the configuration of projects matching your dataset using the new edit button that we've added - instead of moving backwards through the wizard.
Our team set out to build an exceptional football team for less than 100M Euros. Using data provided in the FIFA 2020/2021 dataset - the video game - we built a prediction model in order to find the key performance attributes for each position. Then, we used this to pick out a team of excellent but undervalued players.
Maria and Paul analyzed a dataset of products from a bakery in Edinburgh to discover the associations between menu items. In this video, they walk through the process of conducting a simple product association analysis that could be used for any e-commerce or retail business.
We've been focusing on improving our data exploration capabilities and have added some features making it easier to build projects with big datasets and dive straight into important aspects of your analysis. On top of this, we are working on making Graphext a more powerful data cleaning and preprocessing tool.
Projects in Graphext just got bigger. Now, you can create projects using datasets with hundreds of thousands of rows like this one that Victoriano created using 215 thousand rows of data about salary structures in Spain.
To achieve this, we hide the links between nodes when building larger network visualizations. For the technically minded among us - we moved the storage of network links from JSON into our own database and only draw them for local neighbourhoods.
This means that you can still show connections between a node and its neighbours on larger Graphs. We are really excited about the possibilities that this feature opens up.
- Start from your team's Dataset panel.
- Upload a large dataset.
- Build any type of project using it.
- Start discovering communities inside of your enormous network!
Using the dropdown menu inside of your sidebar variable cards, you can now jump straight into the Compare panel to discern which other variables best explain the difference between values belonging to this variable. Select Open in Compare from the menu list to start understanding your data using compare charts.
We added this feature to make it quicker and simpler for you to jump into a more intricate investigation of the distinguishing features of values in your data.
- Start from your project's Graph, Details or Trends panel.
- Find the variable you want to inspect.
- Click the three dots from the top right of the variable card.
- Choose Open in Compare from the menu list.
- Use the compare charts to pick out the defining features of your values.
We've added the ability to set the type of your variables in more detail. Boolean, Sex and Currency are among the new variable types that you can now make use of in Graphext. From inside your team's Dataset panel, inspect a dataset and use the dropdown under a variable name to set its type to one of the nine options now available.
- Start from your team's Dataset panel.
- Inspect a dataset.
- Click on the dropdown menu underneath a variable name to change its type.
- Choose a new type from the menu list.
- The type of this variable will now update.
We've been delighted with the number of new people using Graphext recently. As a result, we've decided to open up the limit of projects that users can create with a free account. Graphext Public users can now create up to 4 projects.
- Sign up for a Graphext Public account .
- Check out our guides on Getting Started.
- Start analysing your data using Graphext.
- Corrected a problem with clustering configuration in Text → Keyword Co-Occurrence projects.
- Fixed an issue with segment names when performing intersection operations.
- Solved a query text error that was occurring when users searched inside the Graph.
- Added functionality so that longer variable names appear complete rather than incomplete in the Compare panel.
- Fixed issue with dataset vectorization - layout_datset step - as this was occasionally failing on some datasets.
Inspired by an analysis by Ryan Best at FiveThirtyEight, Victoriano and Andy clustered 20 years of Super Bowl commercials. They were interested in which popular brands used characteristics like comedy, sex, patriotism and animals to sell their products. Read More.
Our team have been working on a guide to explain how Graphext can be used to interpret the characteristics, attitudes and preferences of employees. This guide looks at how a prediction model built-in to Graphext might be used to understand why sub-communities of people left their jobs. Read More.