Accurately Detect Fraud with an Akkio Predictive AI Model

If you deal with transactions, inevitably you have to deal with fraud. Finding which transactions are fraudulent so you can intervene can be a complex and overly manual task – but it doesn't have to be.

Detecting fraud can be done automatically without data science resources using Akkio. This tutorial will walk you through using Akkio to upload historical data, build a predictive model, and then deploy it for your business. By the end, you’ll see how Akkio bridges skill gaps with a simple workflow that gives any team the power of intuitive advanced predictive business intelligence to save time and enhance operations.

Gather Historical Data

To start, you’ll need to gather historical data on fraudulent transactions from your database. The data will need to include a column that indicates if the transaction was fraud and any number of potential predictive variables. In this tutorial, we are using fields like number of client transactions, logins in the last hour, and transaction disputes to see which transactions are most likely fraudulent. You can include as many different historical variables as you would like to analyze in aggregate, as long as you have corresponding fraud outputs for each.

If this data is already in a table in your data warehouse like Snowflake, great. If you need to download and aggregate your data in a csv or google sheet (like in the screenshot below), Akkio supports both methods, as long as your data is in tabular form.

‍Important: You must have the column of interest (or, what we’re trying to predict). In this case, the column is named, “is_fraud”.

Import Your Data

Import your data (via table or Akkio-supported integration of choice) in this format.

Ensure All Labels of Data Columns Are Correct 

Akkio will automatically categorize your columns in data types like “number”, “date”, “text”, etc. Before moving to the next step and to ensure that the model we’re building is accurate, double-check that all columns are correctly labeled. For example, “number” should contain only numeric inputs, and columns labeled as “text” should contain only text inputs.

Use “Chat Data Prep” to Prep Your Data

Use the prebuilt Data Cleaning options to quickly perform basic data cleanliness steps, or simply type how you’d like to prepare your data in the box. Akkio will transform your data in seconds, without the need for complex formulas or SQL! Common examples include: 

• “Standardize all date columns” 
• “Remove all null values” 
• “Create a new column that combines columns A and B” 

In this case, the data was already cleaned and pre-processed in Snowflake, which is a native integration to Akkio. If you’re importing your data from a source that needs cleaning, you can easily perform powerful transformations using natural language, or choose from the examples provided.

Create Your Predictive Fraud Detection Model

• Click “Predict” in the top navigation bar

• To train your model, click “Predict” not “Forecast” since we are creating a predictive fraud detection model. For more information on model types, see below or read more here.

• Target the column of interest - in this case, “is_fraud”, in the left-hand “Predict Column”
• If your dataset is live and continues to get updated, Akkio re-evaluates the patterns and can find new patterns of fraud.
• For your first model, we recommend choosing the “Fastest” training mode option. Once you’re comfortable with your model and ready to deploy, choose “High Quality” or “Production Quality.”

• Click the blue “Create Predictive Model” button, and Akkio will begin building your model. 
• This may take a couple of minutes to half an hour, depending on the size of the dataset and training mode. In the background, Akkio is testing a variety of predictive models to find the optimal model for your unique dataset.

Key Insights

The report generated once the model is complete is called the Key Insights report. These insights highlight correlations and driving factors that impact the outcome, which will help you understand where to act next.

• Click “Expand Key Insights”
• Look at Top Fields 
• Click on fields such as “Client Transactions” to view specific insights about which data points were predicted most likely to drive fraud increases. You can see here that transactions that have fewer client transactions (0 to 3) have a positive predictive impact on fraud. This doesn’t mean that every new user is fraudulent, but combined with other data that we have on each transaction, we begin to get a clearer picture of what impacts fraud for our particular dataset. 
• From these insights, we can already start seeing patterns and correlations that we otherwise wouldn’t have noticed with manual data analysis. 

Deploy the model via API 

Now that we’ve built our model and reviewed the new insights, it’s time to deploy it to your application. In the ‘Deploy’ tab, you’ll see a variety of ways that Akkio can deploy your model: as a web app, back to your original data source, or via API. 

For this example, we’ll be deploying the model via API to embed into your company’s platform, enabling you to make calls to the API to detect fraud in seconds when a user inputs information. 

Details on Akkio’s API:

• You must get your API keys and copy them into your API code. Those can be found under the team settings page at the bottom of the Akkio app.
• Once you have installed the Akkio custom package and run the above Usage code you should be able to go to app.akkio.com and see an api test project as shown here.
• The Akkio REST API can be accessed in one of three ways:

1. CURL Commands - Use complete curl commands to negate the need for any library
2. Python - Load the Akkio python library for convenient coding in a Python Environment
3. Node.js - A javascript environment

If you have any questions about our API or other integrations, reach out to support@akkio.com. 

Once you follow the steps above, you can call the API from your platform. The example code below shows how one of our customers calls Akkio’s API for their fraud detection process.

• Store the probability of a transaction being fraudulent, and store it in your database linked to the txn 
• Display the outcome of each transaction (likelihood of fraud) so you can take action – whether that’s flagging your security team, restricting the fraudulent activity, and more.

Once your model is deployed via API or integration of choice and embedded into your product, you’re ready to start detecting transaction fraud. The model will automatically refresh and requires minimal maintenance.

‍

##Conclusion

Congrats! You’ve now built a custom AI model to predict transaction fraud without the need for a data science team. Connect your own data and try this tutorial for yourself with our free two-week trial.

If you have specific questions on how Akkio can transform your fraud detection process, or other AI needs, book an advisory call.