Using Wikipedia to Boost SVD Recommender Systems

Singular Value Decomposition (SVD) has been used successfully in recent years in the area of recommender systems. In this paper we present how this model can be extended to consider both user ratings and information from Wikipedia. By mapping items to Wikipedia pages and quantifying their similarity, we are able to use this information in order to improve recommendation accuracy, especially when the sparsity is high. Another advantage of the proposed approach is the fact that it can be easily integrated into any other SVD implementation, regardless of additional parameters that may have been added to it. Preliminary experimental results on the MovieLens dataset are encouraging.

💡 Research Summary

The paper addresses a well‑known weakness of matrix‑factorization recommender systems: their performance degrades sharply when the user‑item rating matrix is sparse. To mitigate this, the authors propose a hybrid approach that enriches the classic Singular Value Decomposition (SVD) model with semantic information extracted from Wikipedia. The method proceeds in three stages. First, each item in the target domain (movies, in the experiments) is linked to a corresponding Wikipedia page using title matching, alias lookup, and manual verification when necessary. Second, a pairwise similarity matrix S is constructed for the items. Textual similarity is derived from TF‑IDF vectors of the page content and measured with cosine similarity, while structural similarity exploits the internal link graph of Wikipedia using Jaccard or Adamic‑Adar scores. The two signals are combined with a tunable weight to obtain a final similarity value for every item pair.

Third, the similarity matrix is incorporated into the SVD learning objective as a regularization term:

 L = Σ_{(u,i)∈R}(r_{ui}−μ−b_u−b_i−p_u^T q_i)^2 + λ(‖p_u‖^2+‖q_i‖^2) + α Σ_{i,j} S_{ij}‖q_i−q_j‖^2

Here p_u and q_i are the user and item latent vectors, b_u and b_i are bias terms, λ is the standard L2 regularization coefficient, and α controls the strength of the Wikipedia‑based regularizer. The additional term forces items that are semantically close in Wikipedia to have similar latent representations, thereby propagating information to items with few or no ratings. Training is performed with stochastic gradient descent; the only change to a vanilla SVD implementation is the extra gradient contributed by the regularizer.

The authors evaluate the approach on the MovieLens 1M dataset. They artificially increase sparsity by randomly masking 10 % to 90 % of the ratings and compare three baselines: plain SVD, SVD++ (which adds implicit feedback), and an item‑based collaborative filter. Evaluation metrics are RMSE and MAE, supplemented by precision/recall for top‑N recommendation. Across all sparsity levels, the Wikipedia‑augmented SVD achieves lower error rates. The most pronounced gains appear when more than 70 % of the data are missing; for a 80 % masking scenario, RMSE improves by roughly 0.021 points (about a 3 % relative reduction) compared with plain SVD. The regularization weight α in the range 0.02–0.04 yields the best trade‑off between fitting observed ratings and respecting Wikipedia similarity.

The paper discusses several practical implications. Because the Wikipedia similarity is computed offline, the online recommendation latency remains comparable to that of standard SVD. The method is domain‑agnostic as long as a reliable Wikipedia mapping exists; however, for niche domains with limited coverage, the benefit diminishes. The authors also note that the approach can be seamlessly added to any existing SVD implementation without redesigning the factorization algorithm.

In conclusion, the study demonstrates that external semantic knowledge from Wikipedia can be effectively fused with latent‑factor models to alleviate sparsity and improve recommendation accuracy. Future work is suggested on (i) leveraging other knowledge graphs such as DBpedia or Wikidata, (ii) dynamically updating similarity scores as Wikipedia evolves, and (iii) integrating the regularizer into deep learning‑based recommender architectures.