Recommender Systems — Practice

Consider an autoencoder recommender system with one hidden layer. (1 pt) Draw the architecture of the model. (2 pts) Write the equations for each layer, specifying input and output vectors and how ratings are computed. (2 pts) Explain why and under which conditions item-based CF algorithms (e.g. SLIM) can be considered a special case of autoencoder models.

Describe the difference between error metrics, classification metrics, and ranking metrics. Provide an example for each category.

(2 pts) Describe the main idea behind Graph Convolutional Networks and the concept of nodes and edges in recommendation scenarios. (2 pts) Describe the aggregation mechanism used in LightGCN. (2 pts) Describe how LightGCN is trained, including the optimization objective.

Considering a scenario with implicit ratings: (1 pt) Describe the structure of the graph used to represent the data. (1 pt) Describe how to compute the transition probability matrix. (2 pts) Describe how the model is related to item-based collaborative filtering algorithms. (3 pts) Explain the differences between P3, P3alpha, and RP3beta.

(4 pts) Describe the differences between implicit and explicit ratings. Provide an example of an algorithm specifically designed for each of the two types of ratings and explain why. (2 pts) Briefly describe what a multi-stage recommender is and what the purpose of each stage is.

Discuss the main differences between model-based and memory-based recommendation algorithms. Provide an example for each category and explain why it belongs to it.

Consider a two-tower collaborative recommender, each tower a fully-connected neural network with one hidden layer. (1 pt) Draw the architecture of the model. (2 pts) Write the equations for each tower, specifying input and output vectors and how ratings are computed. (2 pts) Explain why and under which conditions traditional ML-based Matrix Factorization can be considered a special case of two-tower models. (1 pt) Explain whether two-tower models are model-based or memory-based.

Consider Bayesian Personalized Ranking (BPR). (2 pts) Describe the main idea and the assumptions it relies upon. (1 pt) Write the loss function using a generic predicted rating. (2 pts) Derive from the loss how to compute the parameter gradients for a matrix-factorization model. (1 pt) Explain how triplets (user, positive item, negative item) are sampled.

Describe how to implement a CF recommender using Factorization Machines (FMs). (1 pt) Describe the structure of the input data used by an FM. (1 pt) Describe the model (equation) used to predict ratings and its parameters. (1 pt) Describe how the parameters of the model are learnt. (2 pts) For explicit ratings, describe the analogy between FMs and Global Effects + Matrix Factorization. (1 pt) Describe how FMs can be used to implement a context-aware recommender system.

Describe how to implement a collaborative recommender for an implicit dataset using a graph-based approach. (1 pt) Describe the structure of the bipartite graph used to represent the interactions. (1 pt) Describe the structure of the incidence (adjacency) matrix. (1 pt) Describe how to compute the probability of jumping from one node to another. (2 pts) Describe how to compute the steady-state probability of reaching a node during a random walk, with and without restart. (1 pt) Describe how to extend the graph to account for side information (item or user attributes).

Discuss the main differences between personalized and non-personalized recommenders. Provide an example of an algorithm for each category and explain why it belongs to it.

Consider a user-based collaborative-filtering recommender system. (2 pts) Describe the main idea and the formula to compute the predictions given the URM and the similarity S, with and without the user bias. (1 pt) Write the formula for cosine similarity. (1 pt) Write the formula for the Pearson correlation similarity and explain when and why it should be used instead of cosine. (1 pt) For implicit ratings and a Top-N task, describe why and how the model equation simplifies. (1 pt) Describe the shrink term — why it improves quality and how it affects the formulas. (1 pt) Describe KNN — why it improves quality and how it affects the formulas.

(2 pts) Describe how autoencoders can be used for recommender systems. (1 pt) Explain the concept of an embedding in the context of autoencoders. (2 pts) Explain how Variational Autoencoders (VAEs) differ from standard autoencoders, and the role of the reparameterization trick.

(2 pts) Describe how BPR exhibits a strong popularity bias and the reason for this. (1 pt) Explain how re-ranking strategies like Maximal Marginal Relevance (MMR) can be used to increase diversity. (2 pts) Describe other re-ranking strategies to diversify recommendations based on item popularity or item similarity.

Describe how to implement a purely content-based recommendation model with machine learning. (2 pts) Write the model in matrix notation and summation notation (using component indices), and include the loss function. (1 pt) Describe the underlying idea and the loss function. (1 pt) Describe which constraints are needed and why. (1 pt) Write the equation to compute predicted ratings in matrix and summation notation.

(2 pts) Describe the main idea behind content-based filtering (CBF) and how it makes recommendations. (1 pt) Explain how TF-IDF can be used to weight item attributes in CBF. (2 pts) Write the model for a content-based filtering recommendation algorithm using matrix notation, and the corresponding loss function.

(2 pts) Explain the concept of denoising autoencoders and how they are trained. (1 pt) Explain what a two-tower model is and how it can be applied to recommender systems. (2 pts) Explain how a shallow autoencoder can be used to obtain an item-item similarity matrix.

(2 pts) Explain how Mean Average Precision (MAP) is calculated for evaluating recommender systems. (1 pt) Explain the purpose of the Average Reciprocal Hit-Rate (ARHR). (2 pts) Explain what the purpose of an A/B test is, and when it is used.

(2 pts) Explain the difference between online and offline evaluation methods for recommender systems. Provide an example of each. (1 pt) Explain the concepts of precision and recall, and how they are calculated. (2 pts) Describe the "Missing As Negative" (MAN) assumption and why it is used in offline evaluation.

(3 pts) Give the definition of precision (P) and recall (R). (4 pts) Plot precision and recall as a function of the number of recommended items K; explain the behaviour of each plot. (2 pts) Plot precision as a function of recall; define Average Precision (AP) and compute it. (3 pts) Exercise — one user with exactly one relevant item. Two algorithms A and B each recommend two items. A puts the relevant item in position 1; B puts it in position 2. Compute precision P and average precision AP for both.

(2 pts) Explain the difference between relevance and diversity in recommender systems and why both are important. (1 pt) Explain how coverage is measured and why it is important. (2 pts) Describe novelty and serendipity as quality indicators.

(2 pts) Explain how Factorization Machines factorize the interaction term to reduce the number of parameters. (1 pt) Describe the equivalence between FM and SVD++ when only collaborative data is used. (2 pts) Explain how implicit ratings are handled in Factorization Machines by creating balanced positive and negative samples.

(2 pts) Explain how global effects are incorporated into the Factorization Machines model. (1 pt) Describe the relationship between Factorization Machines and Matrix Factorization + Global Effects when using explicit ratings. (2 pts) Describe how the imbalance problem is addressed in Factorization Machines with implicit ratings.

(2 pts) Describe the main idea behind Funk SVD and how it differs from full SVD. (1 pt) Explain the assumption on missing interactions in Funk SVD and how it is used. (2 pts) Describe the iterative approach of training Funk SVD using Alternating Least Squares (ALS).

(2 pts) Explain the concept of message passing in Graph Convolutional Networks (GCNs). (1 pt) Describe how LightGCN performs message passing using a simple average aggregation. (2 pts) Explain how predictions are computed using the embeddings generated by LightGCN.

(2 pts) Describe how a user-item interaction graph is constructed for graph-based recommender systems. (1 pt) Explain the concept of a state probability in the context of a random walk on a graph. (2 pts) Describe the PageRank algorithm and how it is used for recommendations.

(2 pts) Explain how random walks are used to find recommendations in a user-item interaction graph. (1 pt) Describe how to compute the state probability for nodes in the graph using a random-walk approach. (2 pts) Explain how to incorporate side information into a graph-based recommender system, including potential problems.

(2 pts) Describe the linear combination approach for hybrid recommender systems, including its advantages and disadvantages. (1 pt) Explain the list combination technique and how it merges results from different models. (2 pts) Describe the concept of co-training and how it is implemented in the S-SLIM hybrid model.

(2 pts) Explain pipelining as a technique for combining different recommendation models. (1 pt) Explain the advantages and disadvantages of pipelining. (2 pts) Discuss why hybrid models are built and the benefits they bring compared to single models.

(2 pts) Describe the main idea behind item-based collaborative filtering and how it computes the predictions. (1 pt) Explain how heuristics like cosine and Pearson can be used to measure item similarities. (2 pts) Explain how the similarity matrix is used to compute predictions in item-based CF.

(2 pts) Explain the idea behind Matrix Factorization (MF) for recommender systems. (1 pt) Describe how parameters are learned in MF using a loss function such as the Mean Squared Error (MSE). (2 pts) Explain how regularization is used in MF to prevent overfitting and its impact on the model.

(2 pts) Describe the basic structure of a neural network layer, including the role of weights, biases, and activation functions. (1 pt) Explain how the Sigmoid and ReLU activation functions work and how they contribute to non-linearity. (2 pts) Explain how Binary Cross Entropy (BCE) can be used as a loss function in a neural network for recommender systems.

(1 pt) Describe the concept of popularity bias and its effect on recommender systems. (2 pts) Discuss the importance of diversity, serendipity and novelty in recommender systems.

(2 pts) Explain how gradient descent is used to optimize model parameters. (1 pt) Write the gradient update equation for the SLIM Elastic-Net model. (2 pts) Describe the iterative process of gradient descent, including sampling, gradient computation, and parameter updates.

(2 pts) Describe how the Mean Squared Error (MSE) is used as a loss function in item-based CF, and the problem of overfitting. (1 pt) Explain why setting the diagonal of the similarity matrix to zero helps avoid trivial solutions and overfitting. (2 pts) Explain how regularization can be used to prevent overfitting, and its effect on the model parameters.

User-based collaborative filtering (8 points). (1 pt) Describe the basic idea (max half a page). (4 pts) Describe different approaches to measure the similarity between users; describe the advantages, disadvantages and peculiarities of each approach. (2 pts) Describe how to estimate ratings using the similarities between users. (1 pt) Write the user-based approach in matrix notation and clearly explain how it is derived.

(2 pts) Describe the difference between user-based and item-based collaborative filtering and their different approaches to generating recommendations. (1 pt) Explain how cosine similarity is calculated and its role in collaborative filtering. (2 pts) Describe the concept of a similarity matrix and how it is used in item-based CF to compute the predictions.