Sparsity-Adaptive Information-Theoretic Co-Clustering (SAITCC) for the Cold-Start Problem on MovieLens-1M Dataset

<jats:p>Recommender systems struggle to help new users who have rated only a few items. This is the cold-start problem, and it matters because new users drive platform growth. The classical ITCC algorithm of Dhillon, Mallela, and Modha (2003) treats all users equally, even though a user with 20 ratings is clearly less reliable than another one with 2,000. In this study, we introduce Sparsity-Adaptive Information-Theoretic Co-Clustering (SAITCC), a method that adjusts for this imbalance by giving less weight to users with sparse rating histories. The proposed method addresses this by augmenting the ITCC objective with a bidirectional maximum-entropy regulariser modulated by Bayesian shrinkage weights w=n/(𝑛+𝜏). The regulariser concentrates on warm entities (reliable) and diminishes on cold ones, preserving cluster-prototype diversity for low-activity users. We derive the SAITCC objective from Lagrangian duality under a maximum-entropy constraint, and construct six variants that are evaluated on the MovieLens-1M dataset across a 1,200-run grid: three unregularised reference points (V0a, V0b, V1) and three regularised ablations (V2-Static, V2-Uni, V2-Bi). The activity-adaptive variants V2-Uni and V2-Bi outperform unregularised ITCC (V1) on cold-start NDCG@10 with Cohen's 𝑑≈1.0 and Holm-Bonferroni-corrected 𝑝&lt;10−5, robust across a 12× perturbation of 𝜏. A cold-start graduation framework distinguishes an absolute cold-start (𝑘visible≤5, no regulariser benefit) from emergent cold-start (𝑘visible≥10, 17–26% lift). The activity-adaptive component is shown to be necessary, since static, uniform-weighted regularisation (V2-Static) underperforms even the unregularised ITCC and exhibits cluster-collapse pathology. Within the ITCC paradigm, V2-Uni uniquely owns the Pareto frontier on joint cold-start NDCG and information preservation. For the investigated dataset, which is a popularity-skewed benchmark, we observed that under top k-NDCG, popularity-based baselines (ItemPop, UserKNN) and BPR-MF outperform all SAITCC variants. This is because our method's scoring function downweights popular items by design. On parsimony and performance grounds, V2-Uni is recommended as the preferred SAITCC variant.</jats:p>