Learning from User generated Data

Question 1

A recommender system creates a list of recommendations based on a user-defined query that expresses the user’s information need. (True/False)

Answer 1

False

A recommender system is an information filtering system that provides a personalized perspective on the available item catalog based on user actions or preferences.

Question 2

An information retrieval system creates a list of documents based on a user-defined query that expresses the user’s information need. (True/False)

Answer 2

True

This is the fundamental function of an information retrieval system.

Question 3

Name the main goals pursued by the course ‘Learning from User-generated Data’ regarding recommender systems.

Answer 3

The main goals are:
* To illustrate approaches to learning from user-generated data
* To provide a sense of how recommender systems are used in real-world applications

Question 4

What are the typical inputs and outputs of recommender systems? Provide examples for each.

Answer 4

Inputs:
* User-item interactions
* User-item ratings
* Personal and item data

Outputs:
* Predicted ratings
* Filtered lists
* Recommendations for the ‘next item’

Question 5

Explain the difference between explicit and implicit user feedback and give an example for each.

Answer 5

Explicit Feedback:
* Direct indication of preference (e.g., ratings)
Implicit Feedback:
* Inferred preferences from interactions (e.g., frequency of consumption)

Question 6

What is the main assumption on which Collaborative Filtering (CF) is based?

Answer 6

The main assumption is that users who had a similar taste in the past will have a similar taste in the future.

Question 7

Describe what an interaction matrix is and why it is a key concept in many recommender systems.

Answer 7

An interaction matrix represents users (rows) and items (columns), with entries showing interactions or ratings. It is key because it forms the basis for many collaborative filtering algorithms.

Question 8

Explain the difference between Memory-based and Model-based Collaborative Filtering.

Answer 8

Memory-based CF:
* Stores all ratings directly
* Predictions are ad-hoc
Model-based CF:
* Factorizes the user-item matrix
* Predictions are based on a learned model

Question 9

Item-based CF scales better for several reasons: (1) item-item similarities can be calculated offline and updated from time to time, and (2) only items that the active user has rated are considered when identifying the nearest neighbors. (True/False)

Answer 9

True

Question 10

Which similarity measures are commonly used in Collaborative Filtering, and why is Adjusted Cosine Similarity relevant for item-based CF?

Answer 10

Common measures:
* Pearson’s correlation coefficient
* Cosine similarity
Adjusted cosine similarity accounts for user-specific rating bias in item-based CF.

Question 11

A major problem with using Truncated SVD for matrix factorization is missing values in the user-item rating matrix. (True/False)

Answer 11

True

Question 12

Model-based CF methods learn exclusively from implicit user feedback, which exacerbates cold-start problems. (True/False)

Answer 12

False

Question 13

When Stochastic Gradient Descent (SGD) is used to create an MF model, a regularization term is used to prevent overfitting. A common choice is Tikhonov regularization. (True/False)

Answer 13

True

Question 14

What are latent factors in matrix factorization, and how are they used to predict ratings?

Answer 14

Latent factors are dimensions derived from rating patterns representing users and items. Predicted ratings are calculated as the inner product of user and item latent vectors.

Question 15

Explain the concept of Singular Value Decomposition (SVD) in the context of matrix factorization for recommender systems.

Answer 15

SVD factorizes a matrix into three smaller matrices: user factors (U), singular values (Σ), and item factors (V^T). It reconstructs the original matrix and enables dimensionality reduction.

Question 16

Consider the following preferences and needs of a user evaluating a new recommender system. What metric corresponds to knowing the precision at the point where precision and recall are equal?

Answer 16

R-precision

Question 17

Two recommender systems have created lists of 3 recommendations for the same user. Both lists achieve the same score for the Reciprocal Rank. (True/False)

Answer 17

True

Question 18

The metrics MRR (Mean Reciprocal Rank) and NDCG (Normalized Discounted Cumulative Gain) consider the position of relevant items in the recommendation list. (True/False)

Answer 18

True

Question 19

Average Precision (AP) considers relevant items that are not in the recommendation list, thus implicitly incorporating recall. (True/False)

Answer 19

True

Question 20

Compared to CG (Cumulative Gain), DCG (Discounted Cumulative Gain) weights the item gains using the position in the recommendation list. (True/False)

Answer 20

True

Question 21

NDCG (Normalized Discounted Cumulative Gain) relates DCG to the ideal DCG to increase the interpretability of the DCG values. (True/False)

Answer 21

True

Question 22

Rank correlation coefficients such as Kendall’s τ or Spearman’s ρ can be used to compare the rankings of 2 (or more) recommender algorithms and determine the extent of their agreement. (True/False)

Answer 22

True

Question 23

What is the range of scores normalized by recommender systems?

Answer 23

0 to 1

This normalization allows for comparisons between different systems and facilitates the interpretation of results.

Question 24

Can rank correlation coefficients like Kendall’s τ or Spearman’s ρ be used to compare recommender algorithms?

Answer 24

True

These coefficients measure the similarity or agreement between rankings of different algorithms.

Question 25

Name and briefly describe the three main scenarios for evaluating recommender systems.

Answer 25

* Offline Evaluation: Based on a fixed dataset, easy to reproduce * Online Evaluation: Conducted with real users, high validity * User Studies: Involves surveys and observations ## Footnote Each scenario has its pros and cons regarding realism, effort, and reproducibility.

Question 26

Explain the difference between Recall@k and Precision@k.

Answer 26

* Precision@k: Proportion of relevant items among the top-k recommendations * Recall@k: Proportion of relevant items recommended out of all available ## Footnote The choice of metric depends on specific use case and priorities.

Question 27

What are 'Beyond-Accuracy Metrics'?

Answer 27

* Diversity * Novelty * Coverage * Serendipity * Explainability ## Footnote These metrics evaluate aspects of recommender systems beyond prediction accuracy.

Question 28

Is content-based filtering well-suited for recommending 'long-tail' items?

Answer 28

True ## Footnote Content features are more objective and less affected by popularity biases.

Question 29

What are lemmatization and stemming used for in text processing?

Answer 29

To reduce the dimensionality of the vector space ## Footnote These techniques reduce the number of unique terms in a vector space model (VSM).

Question 30

True or False: The Vector Space Model (VSM) represents each document as a fixed-dimensional vector of term weights.

Answer 30

True ## Footnote Each dimension corresponds to a unique term from the vocabulary.

Question 31

What is case folding in text preprocessing?

Answer 31

Converts all text characters to lowercase ## Footnote This can introduce semantic ambiguities.

Question 32

What does the normalization assumption in the third monotonicity assumption address?

Answer 32

It is addressed by using cosine similarity ## Footnote Cosine similarity normalizes the vectors to their length.

Question 33

True or False: The logarithmic term in TF and IDF calculations is motivated by Zipf's Law.

Answer 33

True ## Footnote Zipf's Law describes the frequency of word occurrences in natural languages.

Question 34

For short texts, which TF formulation typically outperforms?

Answer 34

Binary TF formulation ## Footnote The logarithmic variant offers no significant advantage for rarely repeated words.

Question 35

What is the Bag of Words (BoW) approach?

Answer 35

A text representation ignoring order, grammar, and syntax ## Footnote Advantages include simplicity and speed; disadvantages include ignoring semantic meanings.

Question 36

What are the essential steps in a text preprocessing pipeline?

Answer 36

* Noise Removal * Lowercasing ## Footnote These steps clean and standardize texts for processing.

Question 37

True or False: LSA is a probabilistic variant of LDA.

Answer 37

False ## Footnote LSA uses Singular Value Decomposition (SVD), while LDA is a probabilistic model.

Question 38

What does LSA perform on the TF-IDF representation of the term-document matrix?

Answer 38

Singular Value Decomposition (SVD) ## Footnote This uncovers latent semantic structures.

Question 39

True or False: The topic modeling approaches pLSA and LDA assume conditional independence of terms and documents.

Answer 39

True ## Footnote This assumption is fundamental to both models.

Question 40

What is the overarching goal of Topic Modeling in text mining?

Answer 40

To describe documents by a set of topics rather than individual words ## Footnote This helps uncover latent semantic structures.

Question 41

True or False: A recommender system that applies graph-based transitivity considers the recommendation task as a graph analysis problem.

Answer 41

True ## Footnote It uses a bipartite graph of users and items with specific path requirements.

Question 42

What is Katz centrality?

Answer 42

A node metric modeling diffusion behavior in networks ## Footnote It measures the number of paths originating from a node, penalizing longer paths.

Question 43

True or False: Graph-based transitivity operates on a bipartite graph with users and items as nodes.

Answer 43

True ## Footnote Edges represent interactions between users and items.

Question 44

What do medial node-based centrality measures consider?

Answer 44

Paths that pass through the node of interest ## Footnote They assess a node's importance based on shortest paths.

Question 45

True or False: K-path centrality and degree centrality are examples of length-based centrality measures.

Answer 45

False ## Footnote They are radial centrality measures, not directly integrated into matrix factorization.

Question 46

What are edge measures in social networks?

Answer 46

They describe local relationships between two nodes ## Footnote Examples include Tie Strength and Edge Betweenness.

Question 47

True or False: In a weighted hybrid system, weights can be learned individually for each user.

Answer 47

True ## Footnote Weights can be adjusted based on available data.

Question 48

True or False: A parallelized design of hybrid recommender systems is an example of early-fusion aggregation.

Answer 48

False ## Footnote It is an example of late-fusion aggregation.

Question 49

What is a parallelized design of hybrid recommender systems an example of?

Answer 49

A late-fusion approach ## Footnote In late-fusion, individual recommenders generate recommendations independently, and the aggregation occurs later.

Question 50

True or False: Monolithic hybridization designs always use late fusion as an aggregation approach.

Answer 50

False ## Footnote Monolithic designs typically employ early-fusion, combining data sources before applying the recommendation algorithm.

Question 51

True or False: Borda rank aggregation is a late-fusion method.

Answer 51

True ## Footnote Borda rank aggregation combines ranks from multiple recommenders, applying the method after individual lists are created.

Question 52

What is CBCF (Content-Boosted Collaborative Filtering) based on?

Answer 52

Enriching real rating data with predictions from a classifier based on content features ## Footnote CBCF combines content-based and collaborative filtering to create a denser matrix of pseudo-ratings.

Question 53

True or False: Self-weighting in Content-Boosted Collaborative Filtering quantifies the algorithm's confidence in the correlation between users' rating vectors.

Answer 53

False ## Footnote Self-weighting quantifies confidence in the content-based prediction used for generating pseudo-ratings.

Question 54

What are the main goals for using hybrid recommender systems?

Answer 54

* To reflect different facets of the items or the domain * To achieve better results (e.g., improved accuracy) * To enable predictions in situations where a single system might not work ## Footnote Hybrid systems leverage strengths of different approaches to overcome individual limitations.

Question 55

List the three main categories of hybrid recommender designs according to Jannach et al.

Answer 55

* Parallelized Design * Monolithic Design * Pipelined Design ## Footnote These categories classify when and how different recommendation approaches are combined.

Question 56

True or False: User context aspects include the target user's activity, mood, and spatio-temporal context.

Answer 56

True ## Footnote User context refers to dynamic aspects of the user's situation during item interaction.

Question 57

True or False: The item context in a multimedia recommender system only includes data encoded in the item's audiovisual signal.

Answer 57

False ## Footnote Item context includes additional data not directly extracted from primary media content, such as tags or album covers.

Question 58

True or False: A music recommender system suggesting more Christmas songs in December is an example of a context-aware recommendation.

Answer 58

True ## Footnote This is an example of considering temporal context.

Question 59

True or False: Context acquisition can be explicit, implicit, or inferential.

Answer 59

True ## Footnote Explicit context is directly collected, implicit derives from behaviors, and inferential uses existing data.

Question 60

What does the Item Purpose describe in a recommender system?

Answer 60

The purpose that the content creator had in mind when creating the item ## Footnote Item Purpose reflects the creator's intention or function of the item.

Question 61

True or False: The User Background describes dynamic information about the user.

Answer 61

False ## Footnote User Background refers to static characteristics, while dynamic needs fall under User Intent or User Context.

Question 62

True or False: The User Intent describes the purpose that the content creator had in mind when creating the item.

Answer 62

False ## Footnote User Intent describes why the user consumes an item, not the creator's purpose.

Question 63

Explain the difference between passive and active user awareness in context-aware recommender systems.

Answer 63

* Passive User Awareness: Captures context but does not change behavior * Active User Awareness: Automatically integrates new context and adjusts recommendations ## Footnote The difference is in whether context changes trigger immediate adjustments.

Question 64

True or False: In a dimensional model of affect, the valence dimension describes the pleasantness of an emotion.

Answer 64

True ## Footnote Valence indicates the quality of emotion, while arousal indicates intensity.

Question 65

True or False: A common approach to integrating personality into a recommendation algorithm is to redefine user similarity based on personality traits.

Answer 65

True ## Footnote This approach helps in cold-start scenarios by enabling similarity estimation with limited interactions.

Question 66

True or False: Mood is longer-lasting and of low intensity compared to emotion.

Answer 66

True ## Footnote Emotions are specific, intense, and short-lived reactions to stimuli.

Question 67

True or False: The assignment of a participant's responses to the BFI-44 personality instrument to the OCEAN model is achieved through harmonic mean weighting.

Answer 67

False ## Footnote The final score is typically a linear combination of responses.

Question 68

True or False: The personality-aware recommender system by Lu and Tintarev, 2018, creates recommendations by weighting items through a linear combination of rank and diversity.

Answer 68

True ## Footnote This method incorporates personality traits to adjust recommendation diversity.

Question 69

True or False: The base-level activation function in the ACT-R model describes the frequency and recency of item exposure.

Answer 69

True ## Footnote It models how frequently and recently used information is more easily retrieved.

Question 70

What are the three main categories of Psychology-informed Recommender Systems (PIRS)?

Answer 70

* Cognition-inspired Recommender Systems * Personality-aware Recommender Systems * Affect-aware Recommender Systems ## Footnote PIRS utilizes psychological insights to refine recommendations.

Question 71

Explain the concept of the Ebbinghaus forgetting curve.

Answer 71

It describes the decline in memory performance over time ## Footnote In recommender systems, it models how user interests lose relevance, suggesting a time-based decay weighting for interactions.

Question 72

True or False: Common strategies for mitigating unwanted biases in a recommender system include regularization, data rebalancing, and adversarial training.

Answer 72

True ## Footnote These strategies address bias through various processing techniques.

Question 73

True or False: Many common recommendation algorithms favor popular items over less popular ones.

Answer 73

True ## Footnote This is known as popularity bias or the 'rich-get-richer' effect.

Question 74

True or False: Individual fairness implies that similar users are treated similarly.

Answer 74

False ## Footnote Individual fairness and group fairness are distinct concepts; one does not guarantee the other.

Question 75

What is popularity bias in recommender systems?

Answer 75

Popularity bias is known as the 'rich-get-richer' effect, where already popular content tends to be recommended more frequently, reinforcing the popularity distribution. ## Footnote This phenomenon can lead to a lack of diversity in recommendations.

Question 76

Individual fairness implies that similar users are treated similarly. True or False?

Answer 76

False ## Footnote Individual fairness and group fairness are different concepts. Achieving group fairness does not guarantee individual fairness for every user.

Question 77

What percentage of songs in a music recommender system's catalog were created by female artists if the recommendation lists contain an average of 50% songs by female artists?

Answer 77

42% ## Footnote The recommender system attempts to mitigate societal bias by striving for a 50% representation of female artists in recommendations.

Question 78

Filtering items created by the majority group of producers from the recommender list is considered a post-processing bias mitigation strategy. True or False?

Answer 78

True ## Footnote Post-processing strategies are techniques applied after the recommendation algorithm has generated its initial results.

Question 79

What is the difference between 'Societal Bias' and 'Statistical Bias' in recommender systems?

Answer 79

* Societal Bias: Discrepancy between the ideal world and reality (e.g., equal representation of genders). * Statistical Bias: Discrepancy between reality and its representation in the system or model. ## Footnote Societal bias concerns external norms, while statistical bias relates to data representation and model building.

Question 80

What types of harm can arise from harmful biases in recommender systems?

Answer 80

* Distributional Harm: Unjust denial of resources or advantages to a person or group. * Representational Harm: Misrepresentation or encoding of stereotypes in the system. ## Footnote These harms impact social and ethical dimensions beyond system performance.