Budget-Constrained Item Cold-Start Handling in Collaborative Filtering Recommenders via Optimal Design

It is well known that collaborative filtering (CF) based recommender systems provide better modeling of users and items associated with considerable rating history. The lack of historical ratings results in the user and the item cold-start problems. The latter is the main focus of this work. Most of the current literature addresses this problem by integrating content-based recommendation techniques to model the new item. However, in many cases such content is not available, and the question arises is whether this problem can be mitigated using CF techniques only. We formalize this problem as an optimization problem: given a new item, a pool of available users, and a budget constraint, select which users to assign with the task of rating the new item in order to minimize the prediction error of our model. We show that the objective function is monotone-supermodular, and propose efficient optimal design based algorithms that attain an approximation to its optimum. Our findings are verified by an empirical study using the Netflix dataset, where the proposed algorithms outperform several baselines for the problem at hand.

💡 Research Summary

The paper tackles the item cold‑start problem in collaborative‑filtering (CF) recommender systems by formulating it as a budget‑constrained user‑selection task. When a new item arrives, the system can ask at most B users from a pool Uᵢ to rate it; the goal is to choose these users so that the expected mean‑squared error (MSE) of rating predictions for all remaining users is minimized. The authors assume that a latent‑factor model (LFM) has already been trained on historical data, providing each user u with a bias bᵤ and a k‑dimensional latent vector Pᵤ, while each item i would have an unknown bias bᵢ and latent vector Qᵢ. Ratings follow the standard LFM equation

  rᵤᵢ = μ + bᵤ + bᵢ + QᵢᵀPᵤ + εᵤᵢ,

where εᵤᵢ is zero‑mean noise (i.i.d. Gaussian or merely independent).

Problem decomposition
The overall problem splits naturally into (A) Rating Prediction – given a set of B observed ratings, estimate (bᵢ, Qᵢ) and predict the rest of the ratings – and (B) User Selection – choose the B users that lead to the most accurate predictions. Under the LFM assumption, the optimal estimator for (bᵢ, Qᵢ) is the ordinary least‑squares (OLS) solution, possibly regularized. The OLS estimator can be written compactly as

  ( b̂ᵢ , Q̂ᵢ ) = (X_BᵀX_B)⁻¹ X_Bᵀ y_B,

where X_B stacks the augmented user vectors (1, Pᵤ)ᵀ for the selected users and y_B contains the bias‑removed observed ratings.

Objective function properties
The expected prediction error after selecting a set S ⊆ Uᵢ can be expressed as a function f(S) that depends only on the information matrix X_SᵀX_S. The authors prove that f is monotone (adding users never hurts) and supermodular (diminishing returns). This structural property guarantees that a greedy algorithm achieves a (1 − 1/e) approximation to the optimal selection, a classic result from submodular optimization.

Optimal design perspective
Instead of directly evaluating f for every subset (which is infeasible), the paper adopts the A‑optimality criterion from optimal experimental design. A‑optimality seeks to minimize the trace of the covariance matrix of the OLS estimator, i.e.,

  trace