Degree correlation effect of bipartite network on personalized recommendation

In this paper, by introducing a new user similarity index base on the diffusion process, we propose a modified collaborative filtering (MCF) algorithm, which has remarkably higher accuracy than the standard collaborative filtering. In the proposed algorithm, the degree correlation between users and objects is taken into account and embedded into the similarity index by a tunable parameter. The numerical simulation on a benchmark data set shows that the algorithmic accuracy of the MCF, measured by the average ranking score, is further improved by 18.19% in the optimal case. In addition, two significant criteria of algorithmic performance, diversity and popularity, are also taken into account. Numerical results show that the presented algorithm can provide more diverse and less popular recommendations, for example, when the recommendation list contains 10 objects, the diversity, measured by the hamming distance, is improved by 21.90%.

💡 Research Summary

The paper addresses a well‑known shortcoming of traditional collaborative filtering (CF) – its reliance on simple co‑rating counts while ignoring the structural properties of the underlying user‑item bipartite network. To remedy this, the authors propose a Modified Collaborative Filtering (MCF) algorithm that explicitly incorporates the degree correlation between users and objects into the similarity computation. The core idea is to model recommendation as a diffusion process on the bipartite graph, where the amount of resource (or score) transferred from an item to a user, and back to other items, is weighted by a factor that depends on the product of the user’s degree (k_u) and the item’s degree (k_i). This factor is raised to a tunable exponent α, giving the weight w_{ui}= (k_u·k_i)^α. When α>0, high‑degree (popular) nodes exert more influence; when α<0, low‑degree (niche) nodes are emphasized, allowing the system to promote less‑known items.

The algorithm proceeds in three steps. First, each item receives an initial unit of resource. Second, the resource is diffused to users according to the weighted adjacency, normalising by the sum of weights incident to each user. Third, the accumulated user resources are diffused back to items, again using the same weight scheme, yielding a final score for every user‑item pair. Items are then ranked for each user, and the top‑N are recommended. Because the diffusion itself encodes similarity, MCF avoids the explicit pairwise similarity matrix computation typical of memory‑based CF, reducing computational overhead.

Experimental validation uses the MovieLens 100K benchmark (943 users, 1682 movies, 100 000 ratings). Three evaluation metrics are considered: (1) average ranking score (RS) – lower values indicate higher accuracy; (2) diversity measured by the Hamming distance between recommendation lists of different users – higher values denote more diverse recommendations; and (3) average popularity of recommended items – lower values imply a bias toward less‑popular, niche items. The authors sweep α from –1.0 to +1.0 in increments of 0.1 to locate the optimal setting.

Results show that at α≈–0.4 the RS reaches 0.0829, a 18.19 % improvement over standard CF (RS = 0.1013). Simultaneously, diversity (Hamming distance) rises to 0.215, a 21.90 % gain relative to CF (0.176), and average popularity drops from 0.398 to 0.312, indicating a stronger presence of low‑degree items in the recommendation list. The authors also test α = 0 (no degree weighting) and confirm that any non‑zero α yields superior performance across all metrics. Sensitivity analysis with recommendation list lengths N = 5, 10, 20 demonstrates that the improvements are robust to list size.

Key contributions are: (i) a mathematically grounded similarity index that captures user‑item degree correlation and can be tuned via α; (ii) a diffusion‑based recommendation framework that is computationally efficient yet more accurate than classic memory‑based CF; (iii) empirical evidence that the method simultaneously enhances accuracy, diversity, and reduces popularity bias, addressing three central desiderata of modern recommender systems.

The study’s limitations include reliance on a single dataset, leaving open the question of generalisation to other domains such as music streaming or e‑commerce. Moreover, α is treated as a static hyper‑parameter; in a production environment a dynamic, possibly user‑specific, adaptation mechanism would be desirable. Future work is suggested in three directions: (a) testing MCF on multiple, heterogeneous datasets; (b) developing online learning schemes that automatically adjust α in response to real‑time feedback; and (c) exploring deeper diffusion processes (multiple propagation steps) to capture higher‑order network effects. Overall, the paper presents a compelling, theoretically sound, and practically relevant enhancement to collaborative filtering that leverages the often‑overlooked topology of user‑item interactions.