Loading…

Worked solutions

IB Computer Science HL · Paper 3

Computer Science, Higher Level

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

Keep your copy of the paper open alongside. We don't reproduce the questions: each part below is labelled to match the November 2023 Global 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

Hardware resources in IaaS platforms and definition of machine learning hyperparameters

Resource management
Modelling and simulation

(a)

2 marks

Identify two core computational resources provisioned through Infrastructure as a Service (IaaS):
1) Virtualized compute instances and central processors: On-demand virtual machines (VMs) allocated with customizable vCPU cores and RAM to execute machine learning models and server-side processing.
2) Scalable distributed block and object storage: Persistent cloud storage buckets and virtual disks capable of holding massive multi-gigabyte media files, user databases, and training datasets.

Mentor tip

Name two core IaaS resources such as virtual machines/processors, primary RAM, block/object storage, or virtual networks.

(b)

2 marks

Define the term hyperparameter in the context of machine learning algorithms:
A hyperparameter is an external configuration variable set by engineers prior to initiating the model training process, rather than a parameter learned directly from training data.
Hyperparameters govern the training dynamics, structure, and behavior of the learning algorithm (for example, the number of neighbors $k$ in k-NN, the regularization constant $λ$ , or the learning rate $η$ in gradient descent).

Mentor tip

Define hyperparameters as external values configured before training begins to control the learning process, giving an example like $k$ in k-NN or learning rate.

Question 2

8 marks

Risks to user anonymity from behavioural telemetry and mitigating popularity bias

System fundamentals
Modelling and simulation

(a)

4 marks

Explain how collecting and analyzing implicit behavioural data to personalize adverts can compromise a user's fundamental right to anonymity:
1) Unobtrusive behavioral fingerprinting: Tracking granular interaction sequences—such as micro-pauses, video scrubbing cadence, specific viewing timestamps, search history, and device headers—builds a unique behavioral profile that distinguishes an individual from other users.
2) De-anonymization via dataset linkage: Even when records lack direct personal names, third-party advertising brokers can cross-reference high-dimensional behavioral patterns with external publicly available datasets (such as social media profiles or transaction registries), unmasking the user's real-world identity.
3) Sensitive attribute inference: Advanced machine learning models can infer private personal attributes (including political leaning, religious views, age bracket, sexual orientation, or medical conditions) purely from viewing habits without the user's explicit consent.
4) Lack of transparent oversight: Because implicit tracking operates silently in the background, users have no direct visual indicator that a unique, deanonymizing profile is being constructed.

Mentor tip

Explain how granular behavioral tracking creates unique profiles that can be cross-referenced with external data to de-anonymize individuals and infer sensitive personal characteristics.

(b)

4 marks

Explain how popularity bias can be mitigated in a media recommender system:
1) Catalog diversification and stochastic exploration: Rather than populating recommendation feeds exclusively with top-trending content, the algorithm injects a controlled percentage of randomly sampled long-tail, niche, or newly published artist uploads into the recommendation queue.
2) Balanced hybrid recommendation architecture: The system combines collaborative filtering with content-based feature matching, ensuring newly uploaded video clips with zero views are recommended based on musical genre, instrument tags, and artistic category rather than raw interaction counts.
3) Re-ranking algorithms with fairness constraints: A post-processing re-ranking layer adjusts raw predictive scores using fairness penalties, discounting items with massive view counts and boosting high-quality niche creators to ensure balanced visibility.
4) Regularization across latent factor models: In matrix factorization, applying penalty constraints to item latent vectors prevents popular items from skewing the latent feature space, maintaining balanced representation for less-frequently rated work.

Mentor tip

Explain strategies such as long-tail catalog diversification, hybrid content-based fallbacks, regularization penalties, and post-processing re-ranking layers.

Question 3

6 marks

Matrix factorization workflow for personalized content recommendation

Modelling and simulation

6 marks

Describe the sequential computational steps required to recommend content using matrix factorization, assuming the initial user-item interaction matrix has been populated:
1) Latent matrix initialization: The algorithm decomposes the original sparse user-item interaction matrix $R$ (of size $m \times n$ ) into two lower-dimensional dense matrices: a user-feature matrix $U$ (size $m \times k$ ) and an item-feature matrix $V$ (size $n \times k$ ), where $k$ represents the number of latent factors. Values in $U$ and $V$ are initialized with small random floating-point numbers or via truncated singular value decomposition (SVD).
2) Iterative model training via optimization: The system calculates the reconstruction error between observed non-empty ratings $R_{u, i}$ and the dot-product estimate $\hat{R}_{u, i} = U_{u} \cdot V_{i}^{T}$ . An optimization algorithm (such as stochastic gradient descent or alternating least squares) iteratively updates the latent vectors to minimize a regularized cost function:
$min_{U, V} \sum_{(u, i) \in R} (R_{u, i} - U_{u} V_{i}^{T})^{2} + λ (∥ U_{u} ∥^{2} + ∥ V_{i} ∥^{2})$
3) Convergence on optimal latent vectors: Parameter updates repeat iteratively until the cost function reaches a minimal threshold or optimization reaches a predefined epoch limit.
4) Full matrix reconstruction and prediction: Multiplying the trained matrices $U$ and $V^{T}$ reconstructs a dense predicted matrix $\hat{R} = U \times V^{T}$ , generating predicted rating scores for every empty, unrated cell.
5) Filtering, ranking, and recommendation generation: For a given target user, previously viewed items are filtered out, unviewed items are ranked in descending order according to their predicted scores, and the highest-ranked items are served as personalized recommendations.

R \approx U \times V^{T}, \overset{r}{^}_{u, i} = u_{u} \cdot v_{i}^{T} = f = 1 \sum k u_{u, f} v_{i, f} L = (u, i) \sum (r_{u, i} - u_{u} v_{i}^{T})^{2} + λ (∥ u_{u} ∥^{2} + ∥ v_{i} ∥^{2})

Mentor tip

Present the logical pipeline: low-rank matrix initialization, iterative loss minimization (gradient descent with regularization), dense dot-product reconstruction, score ranking, and top- $N$ recommendation.

Question 4

12 marks

Evaluation of supervised k-nearest neighbour algorithms for precision and accuracy in recommender systems

Modelling and simulation

12 marks

Discuss whether selecting and implementing a supervised k-nearest neighbour (k-NN) algorithm will produce precise and accurate recommendations for NextStar's platform:
Operational mechanics of k-NN in recommendation:
- Instance-based lazy learning: k-NN is a non-parametric, instance-based model that skips explicit model training; instead, it stores the complete historical interaction dataset and computes similarities across items or users on the fly during inference.
- Geometric neighborhood classification: The algorithm calculates metric distances (using Euclidean, cosine, or Pearson correlation) between the target profile and other catalog entities, identifying the $k$ closest neighbors to forecast preferences.
Factors enhancing recommendation precision and accuracy:
- Interpretability and simplicity: k-NN provides intuitive, transparent logic for recommendations (e.g., 'recommended because you watched similar indie films'), simplifying debugging and user trust.
- High accuracy in dense, well-labeled domains: When catalog items possess clean, multi-dimensional feature labels and small dataset sizes, k-NN calculates precise neighborhood clusters that yield tailored suggestions.
- Tunable hyperparameters: Adjusting the hyperparameter $k$ directly balances precision and variance; small $k$ captures fine-grained sub-genres, while cross-validation can identify the optimal $k$ to avoid overfitting.
Inherent limitations undermining precision and accuracy:
- The curse of dimensionality: As the feature space expands with thousands of media attributes, geometric distances between items converge to near-uniform values in high-dimensional space, degrading neighborhood accuracy.
- Vulnerability to data sparsity and cold starts: In an expanding platform like NextStar, newly uploaded artist videos lack ratings or interaction vectors. Under supervised k-NN, unrated items cannot be placed accurately in the neighborhood space, causing high false-negative rates.
- Computational latency during live inference: Because k-NN computes distances across the entire database at query time ( $O (N \cdot D)$ ), response times increase dramatically as the user base expands, conflicting with low-latency mobile streaming requirements.
- Susceptibility to noisy data and popularity bias: In supervised configurations, items with extensive interactions dominate neighborhood searches, distorting precision for users seeking niche, undiscovered creators.
Synthesis and recommendation:
A standalone supervised k-NN algorithm will achieve high precision on small, well-labeled sub-catalogs, but cannot maintain accuracy, coverage, and real-time performance across NextStar's expanding multimedia catalog due to dimensional sparsity and query latency. NextStar should deploy k-NN as an auxiliary content-based filter for initial item tagging and cold-start onboarding, while implementing a scalable matrix factorization or neural collaborative filtering pipeline for large-scale production recommendations.

sim (u, v) = \frac{u \cdot v}{∥ u ∥∥ v ∥} = \frac{\sum _{i = 1}^{n} u _{i} v _{i}}{\sum _{i = 1}^{n} u _{i}^{2} \sum _{i = 1}^{n} v _{i}^{2}}

Mentor tip

Analyze both sides: strengths of k-NN (simplicity, local accuracy, tunable $k$ ) versus serious limitations (cold start, matrix sparsity, curse of dimensionality, $O (N)$ query latency), concluding with a hybrid recommendation.

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 hyperparameters with model parameters · Q1
Model parameters are learned internal weights derived from training data; hyperparameters are external configuration settings established by engineers before training begins.
02
Assuming matrix factorization requires fully populated matrices · Q3
Matrix factorization specifically targets extremely sparse interaction matrices, learning latent factor representations from the sparse non-zero entries to predict missing values.
03
Overlooking inference latency in lazy-learning algorithms like k-NN · Q4
k-NN performs no offline training; calculating multi-dimensional distances across millions of items during live queries introduces severe runtime latency at scale.

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

Session: November 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 November 2023 Global 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

Hardware resources in IaaS platforms and definition of machine learning hyperparameters

Resource management
Modelling and simulation

(a)

2 marks

Identify two core computational resources provisioned through Infrastructure as a Service (IaaS):
1) Virtualized compute instances and central processors: On-demand virtual machines (VMs) allocated with customizable vCPU cores and RAM to execute machine learning models and server-side processing.
2) Scalable distributed block and object storage: Persistent cloud storage buckets and virtual disks capable of holding massive multi-gigabyte media files, user databases, and training datasets.

Mentor tip

Name two core IaaS resources such as virtual machines/processors, primary RAM, block/object storage, or virtual networks.

(b)

2 marks

Define the term hyperparameter in the context of machine learning algorithms:
A hyperparameter is an external configuration variable set by engineers prior to initiating the model training process, rather than a parameter learned directly from training data.
Hyperparameters govern the training dynamics, structure, and behavior of the learning algorithm (for example, the number of neighbors $k$ in k-NN, the regularization constant $λ$ , or the learning rate $η$ in gradient descent).

Mentor tip

Define hyperparameters as external values configured before training begins to control the learning process, giving an example like $k$ in k-NN or learning rate.

Question 2

8 marks

Risks to user anonymity from behavioural telemetry and mitigating popularity bias

System fundamentals
Modelling and simulation

(a)

4 marks

Explain how collecting and analyzing implicit behavioural data to personalize adverts can compromise a user's fundamental right to anonymity:
1) Unobtrusive behavioral fingerprinting: Tracking granular interaction sequences—such as micro-pauses, video scrubbing cadence, specific viewing timestamps, search history, and device headers—builds a unique behavioral profile that distinguishes an individual from other users.
2) De-anonymization via dataset linkage: Even when records lack direct personal names, third-party advertising brokers can cross-reference high-dimensional behavioral patterns with external publicly available datasets (such as social media profiles or transaction registries), unmasking the user's real-world identity.
3) Sensitive attribute inference: Advanced machine learning models can infer private personal attributes (including political leaning, religious views, age bracket, sexual orientation, or medical conditions) purely from viewing habits without the user's explicit consent.
4) Lack of transparent oversight: Because implicit tracking operates silently in the background, users have no direct visual indicator that a unique, deanonymizing profile is being constructed.

Mentor tip

Explain how granular behavioral tracking creates unique profiles that can be cross-referenced with external data to de-anonymize individuals and infer sensitive personal characteristics.

(b)

4 marks

Explain how popularity bias can be mitigated in a media recommender system:
1) Catalog diversification and stochastic exploration: Rather than populating recommendation feeds exclusively with top-trending content, the algorithm injects a controlled percentage of randomly sampled long-tail, niche, or newly published artist uploads into the recommendation queue.
2) Balanced hybrid recommendation architecture: The system combines collaborative filtering with content-based feature matching, ensuring newly uploaded video clips with zero views are recommended based on musical genre, instrument tags, and artistic category rather than raw interaction counts.
3) Re-ranking algorithms with fairness constraints: A post-processing re-ranking layer adjusts raw predictive scores using fairness penalties, discounting items with massive view counts and boosting high-quality niche creators to ensure balanced visibility.
4) Regularization across latent factor models: In matrix factorization, applying penalty constraints to item latent vectors prevents popular items from skewing the latent feature space, maintaining balanced representation for less-frequently rated work.

Mentor tip

Explain strategies such as long-tail catalog diversification, hybrid content-based fallbacks, regularization penalties, and post-processing re-ranking layers.

Question 3

6 marks

Matrix factorization workflow for personalized content recommendation

Modelling and simulation

6 marks

Describe the sequential computational steps required to recommend content using matrix factorization, assuming the initial user-item interaction matrix has been populated:
1) Latent matrix initialization: The algorithm decomposes the original sparse user-item interaction matrix $R$ (of size $m \times n$ ) into two lower-dimensional dense matrices: a user-feature matrix $U$ (size $m \times k$ ) and an item-feature matrix $V$ (size $n \times k$ ), where $k$ represents the number of latent factors. Values in $U$ and $V$ are initialized with small random floating-point numbers or via truncated singular value decomposition (SVD).
2) Iterative model training via optimization: The system calculates the reconstruction error between observed non-empty ratings $R_{u, i}$ and the dot-product estimate $\hat{R}_{u, i} = U_{u} \cdot V_{i}^{T}$ . An optimization algorithm (such as stochastic gradient descent or alternating least squares) iteratively updates the latent vectors to minimize a regularized cost function:
$min_{U, V} \sum_{(u, i) \in R} (R_{u, i} - U_{u} V_{i}^{T})^{2} + λ (∥ U_{u} ∥^{2} + ∥ V_{i} ∥^{2})$
3) Convergence on optimal latent vectors: Parameter updates repeat iteratively until the cost function reaches a minimal threshold or optimization reaches a predefined epoch limit.
4) Full matrix reconstruction and prediction: Multiplying the trained matrices $U$ and $V^{T}$ reconstructs a dense predicted matrix $\hat{R} = U \times V^{T}$ , generating predicted rating scores for every empty, unrated cell.
5) Filtering, ranking, and recommendation generation: For a given target user, previously viewed items are filtered out, unviewed items are ranked in descending order according to their predicted scores, and the highest-ranked items are served as personalized recommendations.

R \approx U \times V^{T}, \overset{r}{^}_{u, i} = u_{u} \cdot v_{i}^{T} = f = 1 \sum k u_{u, f} v_{i, f} L = (u, i) \sum (r_{u, i} - u_{u} v_{i}^{T})^{2} + λ (∥ u_{u} ∥^{2} + ∥ v_{i} ∥^{2})

Mentor tip

Question 4

12 marks

Evaluation of supervised k-nearest neighbour algorithms for precision and accuracy in recommender systems

Modelling and simulation

12 marks

Discuss whether selecting and implementing a supervised k-nearest neighbour (k-NN) algorithm will produce precise and accurate recommendations for NextStar's platform:
Operational mechanics of k-NN in recommendation:
- Instance-based lazy learning: k-NN is a non-parametric, instance-based model that skips explicit model training; instead, it stores the complete historical interaction dataset and computes similarities across items or users on the fly during inference.
- Geometric neighborhood classification: The algorithm calculates metric distances (using Euclidean, cosine, or Pearson correlation) between the target profile and other catalog entities, identifying the $k$ closest neighbors to forecast preferences.
Factors enhancing recommendation precision and accuracy:
- Interpretability and simplicity: k-NN provides intuitive, transparent logic for recommendations (e.g., 'recommended because you watched similar indie films'), simplifying debugging and user trust.
- High accuracy in dense, well-labeled domains: When catalog items possess clean, multi-dimensional feature labels and small dataset sizes, k-NN calculates precise neighborhood clusters that yield tailored suggestions.
- Tunable hyperparameters: Adjusting the hyperparameter $k$ directly balances precision and variance; small $k$ captures fine-grained sub-genres, while cross-validation can identify the optimal $k$ to avoid overfitting.
Inherent limitations undermining precision and accuracy:
- The curse of dimensionality: As the feature space expands with thousands of media attributes, geometric distances between items converge to near-uniform values in high-dimensional space, degrading neighborhood accuracy.
- Vulnerability to data sparsity and cold starts: In an expanding platform like NextStar, newly uploaded artist videos lack ratings or interaction vectors. Under supervised k-NN, unrated items cannot be placed accurately in the neighborhood space, causing high false-negative rates.
- Computational latency during live inference: Because k-NN computes distances across the entire database at query time ( $O (N \cdot D)$ ), response times increase dramatically as the user base expands, conflicting with low-latency mobile streaming requirements.
- Susceptibility to noisy data and popularity bias: In supervised configurations, items with extensive interactions dominate neighborhood searches, distorting precision for users seeking niche, undiscovered creators.
Synthesis and recommendation:
A standalone supervised k-NN algorithm will achieve high precision on small, well-labeled sub-catalogs, but cannot maintain accuracy, coverage, and real-time performance across NextStar's expanding multimedia catalog due to dimensional sparsity and query latency. NextStar should deploy k-NN as an auxiliary content-based filter for initial item tagging and cold-start onboarding, while implementing a scalable matrix factorization or neural collaborative filtering pipeline for large-scale production recommendations.

sim (u, v) = \frac{u \cdot v}{∥ u ∥∥ v ∥} = \frac{\sum _{i = 1}^{n} u _{i} v _{i}}{\sum _{i = 1}^{n} u _{i}^{2} \sum _{i = 1}^{n} v _{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 hyperparameters with model parameters · Q1
Model parameters are learned internal weights derived from training data; hyperparameters are external configuration settings established by engineers before training begins.
02
Assuming matrix factorization requires fully populated matrices · Q3
Matrix factorization specifically targets extremely sparse interaction matrices, learning latent factor representations from the sparse non-zero entries to predict missing values.
03
Overlooking inference latency in lazy-learning algorithms like k-NN · Q4
k-NN performs no offline training; calculating multi-dimensional distances across millions of items during live queries introduces severe runtime latency at scale.

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