Loading…

Worked solutions

IB Computer Science HL · Paper 3

Computer Science, Higher Level

Timezone

TZ1TZ2

Session: May 2023
Marks: 30
Duration: 1 hour
Calculator: Not allowed
Questions: 4

Q1 Q2 Q3 Q4

Keep your copy of the paper open alongside. We don't reproduce the questions: each part below is labelled to match the May 2023 TZ2 paper, which your school or the IB can provide.

Practice mode

Hide our working until you've had a go.

Questions

Questions 1–4 · 30 marks

Question 1

4 marks

Infrastructure as a Service (IaaS) benefits and stochastic gradient descent optimization

Resource management
Modelling and simulation

(a)

2 marks

Identify two reasons why NextStar would choose Infrastructure as a Service (IaaS) to host their platform:
1) Dynamic elasticity and resource scaling: NextStar can provision virtual compute instances and distributed block storage on demand as user video uploads increase, scaling infrastructure without purchasing on-premises hardware.
2) Fine-grained administrative and architectural control: IaaS grants complete root-level control over the operating system, network firewalls, and custom machine learning libraries (such as TensorFlow or PyTorch), which managed PaaS or SaaS environments often restrict.

Mentor tip

State two concrete business or technical advantages such as on-demand scalability, cost efficiency (pay-as-you-go), or granular control over the software stack.

(b)

2 marks

Define the term stochastic gradient descent in machine learning optimization:
Stochastic gradient descent is an iterative optimization algorithm used to minimize a model's cost or loss function by calculating the gradient of the loss with respect to model parameters using a single randomly selected training sample (or mini-batch) at each step.
It updates model weights incrementally in the direction of steepest descent, achieving faster per-iteration updates and providing the stochastic variance needed to escape shallow local minima in large datasets.

θ_{t + 1} = θ_{t} - η \nabla L_{i} (θ_{t})

Mentor tip

Define SGD as an iterative optimization method updating weights using randomly sampled single data points to minimize the loss function.

Question 2

8 marks

Evaluation of implicit behavioural telemetry and k-nearest neighbour similarity

System fundamentals
Modelling and simulation

(a)

4 marks

Discuss whether the benefits of NextStar collecting and utilizing implicit behavioural data outweigh the associated privacy and ethical concerns:
Benefits of implicit behavioural tracking:
- Continuous, non-intrusive data volume: Users frequently skip explicit rating forms, but implicit signals (such as video watch completion percentage, repeated replays, search queries, and browse clicks) generate abundant, continuous telemetry automatically.
- Higher authenticity: Implicit actions reflect genuine, subconscious preferences and real user interests, avoiding the subjective rating biases or social desirability skew common in explicit surveys.
Concerns and ethical risks:
- Violation of privacy and user autonomy: Silently tracking browsing cadence, session timestamps, and device fingerprints without explicit user awareness infringes on user expectations of privacy and can lead to unintended profiling.
- Data ambiguity and noise: Implicit signals can be misleading; for instance, a user pausing a video to take a phone call may be recorded as high engagement, skewing recommendation models with noisy data.
Synthesis and measured conclusion:
While implicit data is essential for delivering timely recommendations to passive consumers, the associated privacy risks can compromise community trust. NextStar can balance these priorities by adopting clear consent mechanisms, providing transparency regarding collected metrics, and allowing users to clear their watch histories or disable personalized tracking.

Mentor tip

Examine both benefits (abundant data, authentic subconscious user behavior) and concerns (covert tracking, privacy infringement, noisy data), concluding with a balanced recommendation.

(b)

4 marks

Describe how the k-nearest neighbour (k-NN) algorithm uses item-feature similarity to generate recommendations:
1) Multi-dimensional feature vector representation: Each media item in the catalog is represented as a point in an $n$ -dimensional geometric feature space, where axes correspond to quantitative attributes (such as video genre scores, duration, tempo, audio acousticness, and visual tag weights).
2) Metric distance computation: When a target item is viewed, the algorithm computes its mathematical distance (using Euclidean distance, Manhattan distance, or cosine similarity) against all other items in the database: $d (u, v) = \sum_{i = 1}^{n} (u_{i} - v_{i})^{2}$ .
3) Nearest neighbor identification: The algorithm sorts items by ascending distance and isolates the $k$ closest items (the $k$ -nearest neighbors) in the feature space.
4) Ranking and recommendation: Items within this nearest cluster that the user has not yet viewed are ranked by proximity and presented as recommendations.

d (x, y) = i = 1 \sum n (x_{i} - y_{i})^{2}, sim (x, y) = \frac{x \cdot y}{∥ x ∥∥ y ∥}

Mentor tip

Explain the sequential process: representing items as multi-dimensional vectors, calculating geometric distance (e.g. Euclidean or cosine similarity), identifying the $k$ closest items, and recommending the top matches.

Question 3

6 marks

Inherent architectural limitations of collaborative filtering recommender systems

Modelling and simulation

6 marks

Explain two major limitations of collaborative filtering systems in a new media platform:
1) Cold-start dilemma for new users and new items:
- New item challenge: Collaborative filtering relies entirely on historical user interactions (ratings, views, likes). When a new artist uploads a video clip, it possesses zero interaction history, meaning collaborative algorithms cannot compute similarity against other items, leaving new talent unrecommended.
- New user challenge: A newly registered user has logged zero views or ratings, preventing the system from identifying similar user neighbors or predicting preferences.
2) Interaction matrix sparsity and poor computational scalability:
- Matrix sparsity: In platforms with millions of users and videos, an individual user interacts with only a tiny fraction (e.g., $< 0.01%$ ) of the catalog. The resulting user-item interaction matrix is overwhelmingly empty, making it statistically difficult to discover overlapping rating patterns between users.
- Non-linear scalability limits: Calculating pairwise similarities across large matrices requires substantial memory and computational processing that scales non-linearly ( $O (M \cdot N)$ or $O (M^{2})$ ), causing significant latency during live recommendation queries unless compressed through matrix factorization or dimensionality reduction.

Sparsity = 1 - \frac{∣ Interactions ∣}{∣ Users ∣ \times ∣ Items ∣} \approx 99.9%

Mentor tip

Structure your response around two detailed limitations: the cold-start problem (new items and new users) and matrix sparsity/scalability overhead.

Question 4

12 marks

Comprehensive evaluation of recommender system development: algorithms, training data, and metrics

Modelling and simulation

12 marks

Discuss whether the primary challenges in developing an effective recommender system can be resolved through algorithmic choice, training data engineering, and evaluation methodology:
Core development challenges facing NextStar:
- Promoting emerging talent versus popularity bias: Collaborative models favor established videos with high view counts, creating a feedback loop where popular content dominates and undiscovered artists remain hidden.
- Cold-start and data sparsity: New users and uploads lack interaction telemetry.
- Scalability and real-time response: Serving low-latency recommendations across thousands of concurrent mobile streams.
Influence of algorithm selection:
- Strengths and limitations of k-NN: Effective and interpretable on small item-feature spaces, but computationally slow during inference on large databases and vulnerable to the curse of dimensionality.
- Strengths of matrix factorization: Singular value decomposition (SVD) and alternating least squares compress sparse matrices into compact latent factor vectors, uncovering hidden taste dimensions and reducing prediction latency.
- Hybrid architecture solution: Combining content-based filtering (for new items) with collaborative matrix factorization (for mature user pools) addresses the cold-start limitation directly.
Influence of training data quality and preparation:
- Data cleaning and normalization: Pre-processing explicit ratings to account for user rating habits (some users give 4 stars as an average, others give 2) and de-duplicating artist tags ensures balanced inputs.
- Mitigating noise and bias: Filtering out artificial clickbait and balancing datasets using synthetic sampling or weighted penalties prevents popular items from monopolizing model training.
- Incorporating implicit signals: Augmenting sparse explicit ratings with continuous implicit telemetry (completion percentage, replay frequency) enriches the training set.
Influence of evaluation methodologies:
- Beyond simple accuracy metrics: Traditional statistical metrics like Root Mean Squared Error (RMSE) or Mean Absolute Error (MAE) reward safe predictions of popular items but ignore whether users discover new talent.
- Holistic multi-metric evaluation: Incorporating precision, recall, $F_{1}$ -measure, novelty scores (measuring how unexpected recommendations are), and diversity indexes (ensuring variety across creative categories) provides a balanced benchmark for system performance.
- A/B testing in production: Running live randomized user trials allows engineers to measure genuine engagement and discovery rather than relying solely on offline training validation sets.
Synthesis and conclusion:
Algorithmic choice, curated training data, and multi-metric evaluation can resolve the technical barriers of cold-starts, sparsity, and predictive accuracy. However, technical engineering alone cannot fully resolve subjective human challenges, such as shifting aesthetic tastes, cultural nuances, and creator fairness. NextStar must combine a hybrid machine learning pipeline with ongoing user feedback, transparent privacy controls, and periodic editorial curations to maintain an authentic media discovery ecosystem.

RMSE = \frac{1}{∣ T ∣} (u, i) \in T \sum (r_{u i} - \overset{r}{^}_{u i})^{2}

Mentor tip

Systematically analyze all three prompt pillars: algorithm selection (k-NN vs matrix factorization vs hybrid), training data quality/bias, and evaluation metrics (accuracy vs novelty/diversity), reaching a balanced conclusion.

Our analysis

Where the marks are

Each part sorted by the topic it mostly tests. Modelling and simulation carries the most marks on this paper.

Modelling and simulation24 marks · 80%
Q1, Q2, Q3, Q4
System fundamentals4 marks · 13%
Q2
Resource management2 marks · 7%
Q1

Mentor insights

Traps to watch for

01
Confusing stochastic gradient descent with batch gradient descent · Q1
Batch gradient descent computes the gradient over the entire dataset before making one parameter update; stochastic gradient descent updates weights after evaluating individual randomly sampled data points.
02
Attributing the cold-start problem to lack of server compute · Q3
The cold-start dilemma is an inherent data limitation in collaborative filtering (lack of interaction history for new items/users), not a server processing bottleneck.
03
Relying solely on error metrics like RMSE to evaluate a discovery platform · Q4
Low RMSE proves the model predicts known ratings accurately, but it often reinforces popularity bias; discovery platforms must also evaluate novelty, diversity, and coverage.

Solutions and tips are GeorgeTheTutor's own; the questions are not reproduced here. Not affiliated with or endorsed by the International Baccalaureate Organization.

Want someone to go through your paper with you?

Our maths mentors took these exams recently. One-to-one, they'll find where your marks are going and fix it before May.

Find your maths mentor More past papers

Worked solutions

IB Computer Science HL · Paper 3

Computer Science, Higher Level

Timezone

TZ1TZ2

Session: May 2023
Marks: 30
Duration: 1 hour
Calculator: Not allowed
Questions: 4

Q1 Q2 Q3 Q4

Keep your copy of the paper open alongside. We don't reproduce the questions: each part below is labelled to match the May 2023 TZ2 paper, which your school or the IB can provide.

Practice mode

Hide our working until you've had a go.

Questions

Questions 1–4 · 30 marks

Question 1

4 marks

Infrastructure as a Service (IaaS) benefits and stochastic gradient descent optimization

Resource management
Modelling and simulation

(a)

2 marks

Identify two reasons why NextStar would choose Infrastructure as a Service (IaaS) to host their platform:
1) Dynamic elasticity and resource scaling: NextStar can provision virtual compute instances and distributed block storage on demand as user video uploads increase, scaling infrastructure without purchasing on-premises hardware.
2) Fine-grained administrative and architectural control: IaaS grants complete root-level control over the operating system, network firewalls, and custom machine learning libraries (such as TensorFlow or PyTorch), which managed PaaS or SaaS environments often restrict.

Mentor tip

State two concrete business or technical advantages such as on-demand scalability, cost efficiency (pay-as-you-go), or granular control over the software stack.

(b)

2 marks

Define the term stochastic gradient descent in machine learning optimization:
Stochastic gradient descent is an iterative optimization algorithm used to minimize a model's cost or loss function by calculating the gradient of the loss with respect to model parameters using a single randomly selected training sample (or mini-batch) at each step.
It updates model weights incrementally in the direction of steepest descent, achieving faster per-iteration updates and providing the stochastic variance needed to escape shallow local minima in large datasets.

θ_{t + 1} = θ_{t} - η \nabla L_{i} (θ_{t})

Mentor tip

Define SGD as an iterative optimization method updating weights using randomly sampled single data points to minimize the loss function.

Question 2

8 marks

Evaluation of implicit behavioural telemetry and k-nearest neighbour similarity

System fundamentals
Modelling and simulation

(a)

4 marks

Discuss whether the benefits of NextStar collecting and utilizing implicit behavioural data outweigh the associated privacy and ethical concerns:
Benefits of implicit behavioural tracking:
- Continuous, non-intrusive data volume: Users frequently skip explicit rating forms, but implicit signals (such as video watch completion percentage, repeated replays, search queries, and browse clicks) generate abundant, continuous telemetry automatically.
- Higher authenticity: Implicit actions reflect genuine, subconscious preferences and real user interests, avoiding the subjective rating biases or social desirability skew common in explicit surveys.
Concerns and ethical risks:
- Violation of privacy and user autonomy: Silently tracking browsing cadence, session timestamps, and device fingerprints without explicit user awareness infringes on user expectations of privacy and can lead to unintended profiling.
- Data ambiguity and noise: Implicit signals can be misleading; for instance, a user pausing a video to take a phone call may be recorded as high engagement, skewing recommendation models with noisy data.
Synthesis and measured conclusion:
While implicit data is essential for delivering timely recommendations to passive consumers, the associated privacy risks can compromise community trust. NextStar can balance these priorities by adopting clear consent mechanisms, providing transparency regarding collected metrics, and allowing users to clear their watch histories or disable personalized tracking.

Mentor tip

Examine both benefits (abundant data, authentic subconscious user behavior) and concerns (covert tracking, privacy infringement, noisy data), concluding with a balanced recommendation.

(b)

4 marks

Describe how the k-nearest neighbour (k-NN) algorithm uses item-feature similarity to generate recommendations:
1) Multi-dimensional feature vector representation: Each media item in the catalog is represented as a point in an $n$ -dimensional geometric feature space, where axes correspond to quantitative attributes (such as video genre scores, duration, tempo, audio acousticness, and visual tag weights).
2) Metric distance computation: When a target item is viewed, the algorithm computes its mathematical distance (using Euclidean distance, Manhattan distance, or cosine similarity) against all other items in the database: $d (u, v) = \sum_{i = 1}^{n} (u_{i} - v_{i})^{2}$ .
3) Nearest neighbor identification: The algorithm sorts items by ascending distance and isolates the $k$ closest items (the $k$ -nearest neighbors) in the feature space.
4) Ranking and recommendation: Items within this nearest cluster that the user has not yet viewed are ranked by proximity and presented as recommendations.

d (x, y) = i = 1 \sum n (x_{i} - y_{i})^{2}, sim (x, y) = \frac{x \cdot y}{∥ x ∥∥ y ∥}

Mentor tip

Question 3

6 marks

Inherent architectural limitations of collaborative filtering recommender systems

Modelling and simulation

6 marks

Explain two major limitations of collaborative filtering systems in a new media platform:
1) Cold-start dilemma for new users and new items:
- New item challenge: Collaborative filtering relies entirely on historical user interactions (ratings, views, likes). When a new artist uploads a video clip, it possesses zero interaction history, meaning collaborative algorithms cannot compute similarity against other items, leaving new talent unrecommended.
- New user challenge: A newly registered user has logged zero views or ratings, preventing the system from identifying similar user neighbors or predicting preferences.
2) Interaction matrix sparsity and poor computational scalability:
- Matrix sparsity: In platforms with millions of users and videos, an individual user interacts with only a tiny fraction (e.g., $< 0.01%$ ) of the catalog. The resulting user-item interaction matrix is overwhelmingly empty, making it statistically difficult to discover overlapping rating patterns between users.
- Non-linear scalability limits: Calculating pairwise similarities across large matrices requires substantial memory and computational processing that scales non-linearly ( $O (M \cdot N)$ or $O (M^{2})$ ), causing significant latency during live recommendation queries unless compressed through matrix factorization or dimensionality reduction.

Sparsity = 1 - \frac{∣ Interactions ∣}{∣ Users ∣ \times ∣ Items ∣} \approx 99.9%

Mentor tip

Structure your response around two detailed limitations: the cold-start problem (new items and new users) and matrix sparsity/scalability overhead.

Question 4

12 marks

Comprehensive evaluation of recommender system development: algorithms, training data, and metrics

Modelling and simulation

12 marks

Discuss whether the primary challenges in developing an effective recommender system can be resolved through algorithmic choice, training data engineering, and evaluation methodology:
Core development challenges facing NextStar:
- Promoting emerging talent versus popularity bias: Collaborative models favor established videos with high view counts, creating a feedback loop where popular content dominates and undiscovered artists remain hidden.
- Cold-start and data sparsity: New users and uploads lack interaction telemetry.
- Scalability and real-time response: Serving low-latency recommendations across thousands of concurrent mobile streams.
Influence of algorithm selection:
- Strengths and limitations of k-NN: Effective and interpretable on small item-feature spaces, but computationally slow during inference on large databases and vulnerable to the curse of dimensionality.
- Strengths of matrix factorization: Singular value decomposition (SVD) and alternating least squares compress sparse matrices into compact latent factor vectors, uncovering hidden taste dimensions and reducing prediction latency.
- Hybrid architecture solution: Combining content-based filtering (for new items) with collaborative matrix factorization (for mature user pools) addresses the cold-start limitation directly.
Influence of training data quality and preparation:
- Data cleaning and normalization: Pre-processing explicit ratings to account for user rating habits (some users give 4 stars as an average, others give 2) and de-duplicating artist tags ensures balanced inputs.
- Mitigating noise and bias: Filtering out artificial clickbait and balancing datasets using synthetic sampling or weighted penalties prevents popular items from monopolizing model training.
- Incorporating implicit signals: Augmenting sparse explicit ratings with continuous implicit telemetry (completion percentage, replay frequency) enriches the training set.
Influence of evaluation methodologies:
- Beyond simple accuracy metrics: Traditional statistical metrics like Root Mean Squared Error (RMSE) or Mean Absolute Error (MAE) reward safe predictions of popular items but ignore whether users discover new talent.
- Holistic multi-metric evaluation: Incorporating precision, recall, $F_{1}$ -measure, novelty scores (measuring how unexpected recommendations are), and diversity indexes (ensuring variety across creative categories) provides a balanced benchmark for system performance.
- A/B testing in production: Running live randomized user trials allows engineers to measure genuine engagement and discovery rather than relying solely on offline training validation sets.
Synthesis and conclusion:
Algorithmic choice, curated training data, and multi-metric evaluation can resolve the technical barriers of cold-starts, sparsity, and predictive accuracy. However, technical engineering alone cannot fully resolve subjective human challenges, such as shifting aesthetic tastes, cultural nuances, and creator fairness. NextStar must combine a hybrid machine learning pipeline with ongoing user feedback, transparent privacy controls, and periodic editorial curations to maintain an authentic media discovery ecosystem.

RMSE = \frac{1}{∣ T ∣} (u, i) \in T \sum (r_{u i} - \overset{r}{^}_{u i})^{2}

Mentor tip

Our analysis

Where the marks are

Each part sorted by the topic it mostly tests. Modelling and simulation carries the most marks on this paper.

Modelling and simulation24 marks · 80%
Q1, Q2, Q3, Q4
System fundamentals4 marks · 13%
Q2
Resource management2 marks · 7%
Q1

Mentor insights

Traps to watch for

01
Confusing stochastic gradient descent with batch gradient descent · Q1
Batch gradient descent computes the gradient over the entire dataset before making one parameter update; stochastic gradient descent updates weights after evaluating individual randomly sampled data points.
02
Attributing the cold-start problem to lack of server compute · Q3
The cold-start dilemma is an inherent data limitation in collaborative filtering (lack of interaction history for new items/users), not a server processing bottleneck.
03
Relying solely on error metrics like RMSE to evaluate a discovery platform · Q4
Low RMSE proves the model predicts known ratings accurately, but it often reinforces popularity bias; discovery platforms must also evaluate novelty, diversity, and coverage.

Solutions and tips are GeorgeTheTutor's own; the questions are not reproduced here. Not affiliated with or endorsed by the International Baccalaureate Organization.

Want someone to go through your paper with you?

Our maths mentors took these exams recently. One-to-one, they'll find where your marks are going and fix it before May.

Find your maths mentor More past papers