What’s Key Driver Evaluation, and why do you have to care?

Key Driver Evaluation (KDA) identifies the first components that affect modifications in your knowledge, enabling knowledgeable and well timed choices. Think about managing an ice cream store: in case your suppliers’ ice cream costs spike unexpectedly, you’d wish to rapidly pinpoint the explanations. Be it rising milk prices, chocolate shortages, or exterior market components.

Conventional KDA can compute what drove the modifications within the knowledge. And whereas it gives beneficial insights, it usually arrives too late, delaying crucial choices. Why? As a result of KDA historically entails in depth statistical evaluation, which will be resource-intensive and sluggish.

Automation transforms this situation by streamlining the method and bringing KDA into your decision-making a lot quicker by trendy analytical instruments.

Why Deliver KDA Nearer to Your Selections?

Contemplate the ice cream store situation: one morning, your provide of vanilla ice cream spikes by 63%. A guide KDA may reveal—hours and even days later, relying on when it is run (or whether or not somebody remembers to examine the dashboard)—that milk and chocolate costs have surged, leaving you scrambling for options within the meantime.

Automating this course of by real-time alerts ensures you by no means miss essential occasions:

• Webhook triggers when ingredient costs exceed outlined thresholds.
• Quick automated KDA execution identifies crucial drivers inside moments.
• Immediate alerts allow swift actions like sourcing various suppliers or adjusting costs, safeguarding your small business agility.

These techniques can considerably scale back your response instances, permitting you to mitigate dangers and leverage alternatives instantly, moderately than reacting autopsy.

Automating KDA

Automation considerably streamlines the KDA course of. Typically, you don’t have to react to alerts instantly, however you want your solutions by the following morning. Let’s discover how one can set this up utilizing a sensible instance with Python for in a single day jobs:

def get_notifications(self, workspace_id: str) -> record[Notification]:
    params = {
        "workspaceId": workspace_id,
        "measurement": 1000,
    }
    res = requests.get(
        f"{self.host}/api/v1/actions/notifications",
        headers={"Authorization": f"Bearer {self.token}"},
        params=params,
    )
    res.raise_for_status()

    ResponseModel = ResponseEnvelope[list[Notification]]
    parsed = ResponseModel.model_validate(res.json())
    
    return parsed.knowledge

For this instance, I’ve intentionally chosen 1000 because the polling measurement for notifications. In case you might have greater than 1000 notifications on a single workspace every day, you may wish to rethink your alerting guidelines. Otherwise you may tremendously profit from issues like Anomaly Detection, which I contact on within the final part.

This merely retrieves all notifications for a given workspace, permitting you to run KDA selectively throughout the night time. This protects your computation assets and helps you focus solely on related occasions in your knowledge.

Alternatively, it’s also possible to automate the processing of the notifications with webhooks or our PySDK, so that you don’t should ballot them proactively. You’ll be able to simply simply react to them and have your KDA computed as quickly as any outlier in your knowledge is detected.

Automated KDA in GoodData

Whereas we’re at present engaged on built-in Key Driver Evaluation as an inside characteristic, we have already got a working stream that elegantly automates this externally. Let’s take a look on the particulars. If you happen to’d prefer to study extra or wish to attempt to implement it your self, be at liberty to succeed in out!

Each time a configured alert in GoodData is triggered, it initiates the KDA workflow (by a webhook). The workflow operates in a number of steps:

• Information Extraction
• Semantic Mannequin Integration
• Work Separation
• Partial Summarization
• Exterior Drivers
• Closing Summarization

Information Extraction + Semantic Mannequin integration

First, it extracts details about the metric and filters concerned within the alert, together with the worth that triggered the notification, after which it reads the associated semantic fashions utilizing the PySDK.

The evaluation planner then prepares an evaluation plan primarily based on the precedence of dimensions out there within the semantic mannequin. This plan defines which dimensions and combos can be used to research the metric.

Establishing the Work

The evaluation planner then initiates evaluation employees that execute the plan in parallel. Every employee makes use of the plan to question knowledge and carry out its assigned analyses. These analyses produce alerts that the employee evaluates for potential drivers (what drives the change within the knowledge).

Partial Summarization

If any drivers are discovered, they’re handed to LLM, which selects probably the most related ones primarily based on previous consumer suggestions. It additionally generates a abstract, gives suggestions, and checks for exterior occasions that could possibly be associated.

Exterior Drivers

The evaluation employees course of the plan ranging from a very powerful dimension combos and proceed till all combos are analyzed or the allotted evaluation credit are used up. The credit score system is one thing we applied to permit customers to assign a certain quantity of credit to every KDA in an effort to handle the length and price of the evaluation/LLMs.

Closing Summarization

As soon as the analyses are accomplished, a post-processing step organizes the foundation causes right into a hierarchical tree for simpler exploration and understanding of nested drivers. The LLM then generates an government abstract that highlights a very powerful findings.

We’re at present engaged on enhancing KDA utilizing the semantic mannequin of the metrics. It will assist determine root causes primarily based on combos of underlying dimensions and associated metrics. For instance, a decline in ice-cream margins could also be brought on by a rise within the milk value

A Sneak Peek Into the Future

At the moment, there are three very promising applied sciences that we’re experimenting with.

FlexConnect: Enhancing KDA with Exterior APIs

Increasing automated KDA additional, FlexConnect integrates exterior knowledge by APIs, offering further layers of context. Think about an ice cream store’s knowledge prolonged with exterior market traits, client conduct analytics, or world commodity value indexes.

This integration permits deeper insights past inside knowledge limitations. This will make your decision-making course of extra strong and future-proof. As an illustration, connecting to a climate API may proactively predict ingredient value fluctuations primarily based on forecasted agricultural impacts.

Enhanced Anomaly Detection

Built-in machine studying fashions that spotlight important outliers, enhancing signal-to-noise ratios and accuracy. This may imply that you would be able to simply transfer past easy thresholds and/or change gates. Your alerts can keep in mind the seasonality of your knowledge and easily adapt to it.

Chatbot Integration

We’re at present increasing the probabilities for our AI chatbot, which, in fact, contains Key Driver Evaluation. Quickly, with this functionality, the chatbot can assist you arrange alerts for computerized detection of outliers and ship you notifications about them. Additionally, sooner or later, it might advocate you subsequent steps primarily based on KDA.

The output may look one thing like this:

Sensible Software: Ice Cream Store Instance

As an instance, assume your Anomaly Detection detects a value deviation. Instantly:

1. An automatic KDA course of initiates, revealing milk shortages as the first driver.
2. Concurrently, FlexConnect fetches exterior market knowledge, exhibiting a worldwide dairy scarcity as a consequence of climate situations.
3. An AI agent notifies you through on the spot messaging (or e-mail), providing various suppliers or recommending value changes primarily based on historic knowledge.
4. You’ll be able to then chat with this agent and reveal much more info (or ask it to make use of further knowledge) on the anomaly. The agent has the entire context, as he has been briefed even earlier than you knew in regards to the anomaly.

And whereas this may sound like a really distant future, we’re at present experimenting with every of those! Don’t fear, when every of those options is nearing deployment, we’ll share the PoC with you on this.

Need to study extra?

If you happen to’d prefer to dig deeper into automation in analytics, try our article on how you can successfully make the most of Scheduled Exports & Information Exports. It explores how you can use automation to arrange alerts accurately, in order that they’re helpful and never merely a distraction.

Keep tuned when you’re all in favour of studying extra about KDA, as we’ll quickly observe up with a extra in-depth article whereas additionally exploring its sensible utility in analytics.

Have questions or wish to implement automated KDA in your workflow? Attain out — we’re right here to assist!