Skip to main content

Data Scientist

Real Workflow​

Data science scenarios aren't about "letting AI find insights"—they're about when teams already have data and competing hypotheses, using analysis to narrow down to verifiable directions.

DimensionReal Situation
Trigger PointUser churn rising, conversion declining, A/B experiment results unclear, operational strategy needs review
Existing MaterialsBehavior logs, user profiles, channel sources, version changes, experiment groups and business hypotheses
Pain PointReasons behind metric changes unclear; different teams have different explanations; lack unified analysis framework
DesireCore InterventionData Analyst does funnels, segmentation, anomaly detection and predictive model first draft, organizes verifiable hypotheses
Acceptance ResultTeam gets prioritized recommendations, experiment suggestions and monitoring metrics, then decides product or operational actions

Data Analyst — Suitable for statistical analysis, predictive models, user segmentation, anomaly detection and operational monitoring dashboards.

Complete Conversation Example​

You: Analyze our App's user behavior data and find the reasons for user churn.
    [Upload user_behavior.json]

Data Analyst: Conducting user behavior analysis...

        User Overview
        - Total users: 50,000
        - Monthly Active Users (MAU): 32,000 (64%)
        - Churned users this month: 3,200 (6.4%)

        User Journey Funnel
        Registration → First Use → Activation → Retention → Payment
        100%    78%       45%    32%    8%

        Key Churn Nodes
        Registration → First Use: 22% churn (industry average 15%)
        First Use → Activation: 42% churn (industry average 35%)

        Churn Reason Analysis
        1. Onboarding process too long (35% of users mentioned)
           Average completion time 8 minutes, suggest shortening to 3-5 minutes
        2. Core feature entry not obvious (28% of users mentioned)
           68% of users stay on homepage for over 30 seconds without action
        3. No value felt on first use (22% of users mentioned)
           72% exit without experiencing core features

        User Segmentation
        High-value users (8%): 30+ minutes daily usage
        Potential users (24%): 3+ times weekly but not paid
        Churn risk users (18%): Not logged in for 7 consecutive days

        Prediction
        Without action: Churn rate rises to 8.5% next month
        Optimize onboarding: Expected to reduce churn rate by 2-3%

You: Can you give a priority-ranked improvement plan?

Data Analyst: Improvement Priority Ranking (by expected effect):
        P0: Shorten onboarding process (largest impact)
        P1: Add core feature guidance on homepage
        P2: Design first-use reward mechanism
        P3: Push recall content to churn risk users
        ...

Key Steps​

  1. Define Business Problem — First clarify whether it's churn, conversion, retention, payment or activity problem
  2. Organize Event Definitions — Confirm user ID, event names, timestamps, channels, versions and experiment groups are consistent
  3. Build Funnels and Segmentation — Break down changes by new/returning users, channels, devices, versions and behavior depth
  4. Form Testable Hypotheses — Match hypotheses like "onboarding too long" or "entry not obvious" with data evidence
  5. Build Models and Monitoring — Models only serve as risk ranking; ultimately land on monitorable metrics and experiment plans
  6. Output Action Priority — Sort by impact scope, implementation cost, verification cycle and risk

Final Deliverable​

A user behavior analysis report containing data definitions, funnel analysis, churn attribution, user segmentation, predictive model first draft, monitoring metrics and prioritized experiment suggestions.

Best Practices
  • The more complete the data, the more accurate the analysis. Ideally include user behavior events, timestamps, device information and other dimensions
  • Start with global analysis to understand the overall situation, then dive deep into specific issues. Avoid getting into details from the start
  • Analysis results should translate into executable product improvements—have the agent prioritize by impact scope, implementation cost and verification method
  • User behavior analysis shouldn't jump directly to "reasons"—first confirm tracking definitions, version changes and sample composition are stable