Dynamic recommender system : using cluster-based biases to improve the accuracy of the predictions

It is today accepted that matrix factorization models allow a high quality of rating prediction in recommender systems. However, a major drawback of matrix factorization is its static nature that results in a progressive declining of the accuracy of the predictions after each factorization. This is due to the fact that the new obtained ratings are not taken into account until a new factorization is computed, which can not be done very often because of the high cost of matrix factorization. In this paper, aiming at improving the accuracy of recommender systems, we propose a cluster-based matrix factorization technique that enables online integration of new ratings. Thus, we significantly enhance the obtained predictions between two matrix factorizations. We use finer-grained user biases by clustering similar items into groups, and allocating in these groups a bias to each user. The experiments we did on large datasets demonstrated the efficiency of our approach.

💡 Research Summary

The paper addresses a fundamental limitation of conventional matrix‑factorization (MF) recommender systems: once the latent factor matrices are learned, new user‑item ratings are not incorporated until the next costly re‑factorization, causing a gradual degradation in prediction accuracy. While global, user‑specific, and item‑specific bias terms mitigate some systematic errors, they remain coarse‑grained because each bias is applied uniformly across all items. To overcome this, the authors propose a cluster‑based bias extension that introduces finer‑grained, per‑user bias parameters for groups of similar items.

First, items are clustered using similarity measures derived from content metadata, collaborative similarity, or cosine similarity of rating vectors. The number of clusters K is tuned via cross‑validation. For each cluster c, a distinct bias b_{u,c} is learned for every user u, in addition to the traditional global bias (μ), user bias (b_u), item bias (b_i), and latent vectors (p_u, q_i). The prediction formula becomes:
(\hat{r}{ui}= μ + b_u + b_i + b{u,c(i)} + p_u^\top q_i),
where c(i) denotes the cluster containing item i. Crucially, after an initial offline MF step that computes P and Q, the system updates only the cluster‑specific biases online whenever a new rating arrives. This online update uses stochastic gradient descent on the bias term alone, requiring O(1) computation per rating and no re‑factorization of the full matrix.

The authors analyze computational complexity, showing that the upfront factorization costs O(|R|·d) (|R| = number of observed ratings, d = latent dimension), while subsequent online updates are constant‑time, enabling near‑real‑time adaptation in large‑scale streaming environments. Memory overhead is modest, limited to storing the additional bias matrix of size (number of users × K).

Empirical evaluation on the Netflix Prize dataset and the MovieLens 20M dataset demonstrates the method’s effectiveness. With K ranging from 50 to 200, the cluster‑based model reduces RMSE by 3–5 % relative to a baseline MF model that includes only global/user/item biases. The improvement is especially pronounced in scenarios with frequent introduction of new items, where the baseline’s predictions deteriorate sharply between full re‑factorizations. Moreover, the approach achieves these gains without any additional full matrix factorization, thereby cutting the re‑training interval from weeks to days.

The paper also discusses limitations. The quality of the clustering directly influences performance; overly fine clusters lead to sparse bias parameters and potential over‑fitting, while overly coarse clusters diminish the benefit of per‑cluster bias. The current implementation assigns each item to a single cluster, which may be insufficient for multi‑genre items. Future work is suggested on dynamic clustering, multi‑cluster assignments, and incorporating temporal dynamics into the bias framework.

In summary, by augmenting traditional MF with cluster‑specific user biases that can be updated online, the authors provide a practical solution to the static nature of MF recommender systems. Their method delivers higher prediction accuracy between full factorization cycles while maintaining computational efficiency, making it well‑suited for large‑scale, real‑time recommendation platforms.