Exercise 9: Big data and neural networks

In this notebook, we cover:

Exercise part Time (min)
Part 1: Backpropagation step by step 25 min
Part 2: Interactive neural networks 20 min
Part 3: Code understanding: PyTorch model objects 10 min
Part 4: Use cases quiz: NovaStream 25 min
Wrap-up 5 min
Total 90 min



Part 1: Backpropagation step by step

Go to http://hmkcode.github.io/ai/backpropagation-step-by-step and study the backpropagation algorithm in one iteration.

Note that for simplification, the example uses no activation function.

Part 2: Interactive neural networks

Go to the Tensorflow Playground and experiment with classification problems. Your goal is to achieve a low test loss with the simplest possible model.

In the TensorFlow Playground, select each of the following datasets on the left-hand side:

  1. circle dataset
  2. two-clusters dataset
  3. spiral dataset

For each dataset, modify the neural network design by changing:

  • number of hidden layers
  • number of neurons per layer
  • activation function
  • input features, where useful

Your goal is to find a neural network design that predicts the selected dataset pattern accurately, as indicated by a low test loss.

Do not simply add as many neurons and layers as possible. Instead, try to find a minimal setup that works well.

For each dataset, document:

  • selected input features
  • number of hidden layers
  • number of neurons per layer
  • activation function
  • final test loss

Briefly explain:

  • which settings worked well
  • why the dataset required a simpler or more complex neural network design
  • how the architecture affected the model’s ability to capture the non-linear pattern

Part 3: Code understanding — PyTorch model objects

In sklearn, we often specify the model through parameters such as strings or tuples:

MLPClassifier(
    hidden_layer_sizes=(32, 16),
    activation="relu",
    solver="adam"
)

In PyTorch, we usually define a model as an object. The model inherits functionality from nn.Module. Inspect the following model:

import torch.nn as nn

class MLP(nn.Module):
    def __init__(self, input_dim, hidden_layers, output_dim):
        super().__init__()

        layers = []
        prev_dim = input_dim

        for h in hidden_layers:
            layers.append(nn.Linear(prev_dim, h))
            layers.append(nn.ReLU())
            prev_dim = h

        layers.append(nn.Linear(prev_dim, output_dim))

        self.model = nn.Sequential(*layers)

    def forward(self, x):
        return self.model(x)


model = MLP(
    input_dim=10,
    hidden_layers=[8, 4],
    output_dim=1
)

Write down the sequence of layers created by this model.

Use this format:

Input: 10 features
Linear: 10 -> ...
ReLU
Linear: ... -> ...
ReLU
Linear: ... -> 1
Output: 1 value

Then answer briefly:

  1. How many hidden layers does the model have?
  2. How many trainable Linear layers does it have?
  3. What would change if hidden_layers=[32, 16, 8]?

Part 4: Use Cases Quiz — NovaStream

Background

NovaStream is a rapidly growing digital streaming startup offering movies, series, podcasts, music, and livestream content through a subscription-based platform. Due to strong recent growth, the company plans to scale its operations, improve personalization, optimize infrastructure usage, and strengthen platform security.

Management believes that data-driven and AI-supported decision-making will be critical for future growth. As a result, the company faces a variety of analytical questions that must be addressed using suitable analytical modeling approaches.

Task

For each use case below:

  • Select a suitable analytical modeling approach covered in the lectures.

  • Justify your choice with reference to:

    • the nature of the available input data,
    • the expected targets or outputs,
    • and the suitability of the selected analytical model for the task.

Briefly explain why your chosen approach is appropriate compared to possible alternatives.

Some use cases may not fit perfectly into the standard analytical models discussed in class. In such cases, explain the limitations of standard approaches and discuss which additional analytical or optimization methods may be more appropriate.

# Use Case Recommendation
1 NovaStream wants to identify subscribers who are likely to cancel their subscription within the next 30 days. Available data includes viewing behavior, subscription history, customer support interactions, app usage patterns, and payment information. Early identification would allow the company to proactively offer incentives and reduce customer churn.
2 To improve user engagement, NovaStream wants to recommend movies, series, podcasts, and songs that individual users are likely to consume next. The platform collects information about viewing history, ratings, watch duration, search queries, and similarities between users.
3 NovaStream plans to expand internationally and wants to automatically generate subtitles for livestreams and recorded content in multiple languages. The company wants the generated subtitles to adapt well to conversational language and different speaking styles.
4 The startup has noticed suspicious activity involving accounts that appear to be shared simultaneously across many devices and geographic regions. NovaStream wants to automatically detect unusual behavior patterns that may indicate fraudulent usage or unauthorized account sharing.
5 To accelerate growth, NovaStream plans to dynamically adjust subscription prices and promotional discounts depending on customer demand, competitor actions, viewing behavior, and seasonal effects. The company wants to continuously optimize pricing decisions over time.
# Use Case Recommendation
6 Users can upload custom cover images for podcasts and playlists. NovaStream wants to automatically classify uploaded images into categories such as sports, entertainment, education, politics, or technology in order to improve search and content discovery.
7 NovaStream is introducing a same-day merchandise delivery service for selected cities. The company wants to determine optimal delivery routes for drivers while minimizing transportation costs and ensuring that promised delivery times are met despite changing traffic conditions.
8 NovaStream wants to predict which song, video, or podcast episode a user is most likely to consume next based on recent interaction sequences, time of day, device usage, and historical consumption patterns.
9 Product managers want to understand which combinations of movies, series, podcasts, or songs are frequently consumed together in order to improve homepage design, bundled recommendations, and promotional campaigns.
10 During livestream events, NovaStream offers real-time chat functionality. To maintain a safe platform environment, the company wants to automatically detect toxic, hateful, or abusive chat messages as they are posted.
# Use Case Recommendation
11 The marketing department wants to identify groups of customers with similar viewing and listening behavior in order to design more targeted campaigns and personalized subscription offers. The company does not currently have predefined customer categories.
12 Due to rapid growth, NovaStream’s infrastructure team must better predict future server demand and bandwidth consumption in order to avoid outages during peak usage periods while also minimizing unnecessary cloud costs. Historical usage data across days, weeks, and special events is available.

Wrap-up

In this exercise, we explored how neural networks learn and how analytical models can be selected for different business problems.

Topic Key takeaway
Backpropagation Neural networks learn by gradually adjusting weights based on errors
TensorFlow Playground Different neural architectures vary in how well they model non-linear patterns
PyTorch models Neural networks can be defined as sequences of layers and activation functions
Analytical use cases Different problems require different analytical approaches

In the last sessions, we covered a range of models for tabular and unstructured data in regression, classification, and prediction tasks. At the same time, other problems may require different models, such as time-series forecasting or route optimization.

TipSession 9 survey

Before you wrap up, please complete the Session 9 survey here: ?meta:surveys.session_09.url. Thank you 🙏