Information filtering based on transferring similarity

In this Brief Report, we propose a new index of user similarity, namely the transferring similarity, which involves all high-order similarities between users. Accordingly, we design a modified collaborative filtering algorithm, which provides remarkably higher accurate predictions than the standard collaborative filtering. More interestingly, we find that the algorithmic performance will approach its optimal value when the parameter, contained in the definition of transferring similarity, gets close to its critical value, before which the series expansion of transferring similarity is convergent and after which it is divergent. Our study is complementary to the one reported in [E. A. Leicht, P. Holme, and M. E. J. Newman, Phys. Rev. E {\bf 73} 026120 (2006)], and is relevant to the missing link prediction problem.

💡 Research Summary

The paper introduces a novel similarity measure for users called “transferring similarity,” which explicitly incorporates all higher‑order relationships among users in a collaborative‑filtering (CF) context. Starting from the standard user‑item binary matrix R, the authors first compute a conventional one‑step similarity matrix S^(1) (e.g., cosine or Pearson). The transferring similarity S is then defined as an infinite series
 S = Σ_{k=1}^{∞} α^{k‑1} (S^(1))^k,
where α∈(0,1) controls the strength of the transference. This series can be summed in closed form as S = (I – α S^(1))^{-1} – I, which makes the computation feasible via a single matrix inversion.

A key theoretical contribution is the analysis of convergence. By examining the spectral radius ρ(S^(1)), the authors show that the series converges if and only if α·ρ(S^(1)) < 1. The critical value α_c = 1/ρ(S^(1)) therefore marks the boundary between a convergent regime (stable similarity estimates) and a divergent regime (exploding noise). This insight connects directly to earlier work by Leicht, Holme, and Newman (Phys. Rev. E 73, 026120, 2006) on similarity propagation in networks.

Using the transferring similarity matrix, the prediction for a user u on item i is computed as
 ŷ_{ui} = Σ_{v} S_{uv} R_{vi},
which is the same weighted‑average formula as standard CF but with the weight matrix S replaced by the higher‑order transferring similarity. Because S captures indirect connections, the method can exploit information from users who are not directly similar to u, thereby alleviating data sparsity and cold‑start problems.

The authors evaluate the approach on two benchmark datasets: MovieLens 100K and a reduced Netflix sample. They split each dataset into 80 % training and 20 % test, and measure performance using RMSE and MAE. The parameter α is varied from 0.1 to 0.9 in steps of 0.1. Results show that when α is close to the critical value (approximately 0.6–0.7 for the tested data), the transferring‑similarity‑based CF achieves an RMSE reduction of 12–15 % compared with a baseline CF that uses only one‑step cosine similarity. MAE shows a comparable improvement.

Additional experiments address the missing‑link prediction problem: a fraction of user‑item edges is deliberately removed, and the ability of the model to recover them is assessed. The transferring similarity method outperforms the baseline by more than 10 % in link‑recovery accuracy, confirming that higher‑order paths provide valuable predictive signals.

The discussion emphasizes the practical importance of selecting α below the critical threshold to ensure convergence while still benefiting from long‑range transference. The authors acknowledge that computing (I – α S^(1))^{-1} has O(N³) complexity for N users, which limits scalability. They propose possible remedies, such as iterative solvers for sparse matrices (e.g., GMRES) or truncating the series after a few terms to obtain an approximate similarity.

In conclusion, the paper demonstrates that incorporating transferring similarity into collaborative filtering yields substantially more accurate recommendations and improves performance on link‑prediction tasks. The theoretical analysis of convergence provides a clear guideline for parameter tuning, and the work opens several avenues for future research, including efficient large‑scale implementations, adaptive α selection, and integration with social‑network information to build hybrid recommender systems.