Scenario A: Policy non-renewal

Business requirement:

Predict whether an active policy will fail to renew.

Dataset:

CustomerAge Premium Claims Complaints DidNotRenew
32 12,500 0 0 No
51 22,000 3 4 Yes
44 16,500 1 1 No

Decision:

Known historical label? Yes
Label type?              Yes/No
Correct task?            Supervised binary classification

Possible ML.NET structure:

public sealed class PolicyRenewalInput
{
    public float CustomerAge { get; set; }
    public float PremiumAmount { get; set; }
    public float ClaimCount { get; set; }
    public float ComplaintCount { get; set; }

    public bool DidNotRenew { get; set; }
}

Scenario B: Support-ticket routing

Requirement:

        Predict which department should handle a new ticket.

Possible labels:

Billing
Technical
Sales
General

Decision:

Known historical label? Yes
Label type?              One of four categories
Correct task?            Supervised multiclass classification

C# representation:

 
public sealed class SupportTicketInput
{
    public string TicketText { get; set; } = string.Empty;
    public string Department { get; set; } = string.Empty;
}

Scenario C: Predict resolution time

Requirement:

Estimate how many minutes a support ticket will require.

Historical label:

ResolutionMinutes = 47

Decision:

Known historical label? Yes
Label type?              Numeric quantity
Correct task?            Supervised regression

C# representation:

public sealed class TicketResolutionInput
{
    public float PreviousCustomerTickets { get; set; }
    public float AssignedEngineerExperience { get; set; }
    public float TicketComplexityScore { get; set; }

    public float ResolutionMinutes { get; set; }
}

Scenario D: Discover insurance customer groups

Requirement:

We do not have customer categories. Discover customers with similar behaviour.

Available data-

Premium amount
Claim count
Complaint count
Renewal history
Payment-delay count
Policy count
 Decision:
Known label?   No
Required goal? Discover similar groups
Correct task?  Unsupervised clustering
 Possible ML.NET input:
 

 

public sealed class CustomerBehaviour
{
    public float PremiumAmount { get; set; }
    public float ClaimCount { get; set; }
    public float ComplaintCount { get; set; }
    public float RenewalCount { get; set; }
    public float PaymentDelayCount { get; set; }
}

A clustering pipeline may later look like:

using Microsoft.ML;

var mlContext = new MLContext(seed: 42);

var pipeline = mlContext.Transforms
    .Concatenate(
        "Features",
        nameof(CustomerBehaviour.PremiumAmount),
        nameof(CustomerBehaviour.ClaimCount),
        nameof(CustomerBehaviour.ComplaintCount),
        nameof(CustomerBehaviour.RenewalCount),
        nameof(CustomerBehaviour.PaymentDelayCount))
    .Append(
        mlContext.Clustering.Trainers.KMeans(
            featureColumnName: "Features",
            numberOfClusters: 3));

New term:

K-Means is a clustering algorithm that attempts to organize records around a chosen number of group centres.

Here:

numberOfClusters: 3

means we ask it to find three groups. It does not prove that three is the perfect business answer.

We will train and evaluate clustering models in a later practical module.

Scenario E: Adaptive Assessment OS

Requirement:

Select the next exam-practice question based on the candidate’s changing skill level.

Possible RL design:

RL element Assessment example
Agent Adaptive-question engine
Environment Candidate and assessment session
State Topic mastery, recent answers, time spent
Action Select the next question
Reward Measured learning improvement
Policy Learned question-selection strategy

But this has substantial risks:

  • An experimental policy could disadvantage candidates.
  • Exam marks are not necessarily a valid learning reward.
  • Different candidates may receive unfair experiences.
  • Production exploration may be unacceptable.
  • Regulatory assessment rules may require deterministic delivery.

Therefore, for formal examinations:

Approved deterministic exam rules
              +
Supervised risk or performance models
              +
Human oversight

may be safer than unrestricted reinforcement learning.

RL could first be explored in a non-certification practice simulator.

Build a reusable C# task selector

This code documents the reasoning before algorithm selection:

public enum LearningApproach
{
    DeterministicSoftware,
    SupervisedClassification,
    SupervisedRegression,
    UnsupervisedClustering,
    ReinforcementLearning
}

public sealed record ProblemProfile(
    bool HasHistoricalLabel,
    bool LabelIsCategory,
    bool LabelIsNumericQuantity,
    bool NeedsHiddenGroups,
    bool HasSequentialActionsAndRewards);

public static class LearningApproachSelector
{
    public static LearningApproach Select(ProblemProfile problem)
    {
        if (problem.HasSequentialActionsAndRewards)
            return LearningApproach.ReinforcementLearning;

        if (problem.HasHistoricalLabel && problem.LabelIsCategory)
            return LearningApproach.SupervisedClassification;

        if (problem.HasHistoricalLabel &&
            problem.LabelIsNumericQuantity)
            return LearningApproach.SupervisedRegression;

        if (!problem.HasHistoricalLabel && problem.NeedsHiddenGroups)
            return LearningApproach.UnsupervisedClustering;

        return LearningApproach.DeterministicSoftware;
    }
}

Usage:

var renewalProblem = new ProblemProfile(
    HasHistoricalLabel: true,
    LabelIsCategory: true,
    LabelIsNumericQuantity: false,
    NeedsHiddenGroups: false,
    HasSequentialActionsAndRewards: false);

var result = LearningApproachSelector.Select(renewalProblem);

Console.WriteLine(result);

Expected output:

SupervisedClassification

This selector is a learning aid, not a universal production decision engine. Real selection also depends on data quality, risk, cost and business action.

PreviousNext