Improving Collaborative Filtering based Recommenders using Topic Modelling

Standard Collaborative Filtering (CF) algorithms make use of interactions between users and items in the form of implicit or explicit ratings alone for generating recommendations. Similarity among users or items is calculated purely based on rating overlap in this case,without considering explicit properties of users or items involved, limiting their applicability in domains with very sparse rating spaces. In many domains such as movies, news or electronic commerce recommenders, considerable contextual data in text form describing item properties is available along with the rating data, which could be utilized to improve recommendation quality.In this paper, we propose a novel approach to improve standard CF based recommenders by utilizing latent Dirichlet allocation (LDA) to learn latent properties of items, expressed in terms of topic proportions, derived from their textual description. We infer user’s topic preferences or persona in the same latent space,based on her historical ratings. While computing similarity between users, we make use of a combined similarity measure involving rating overlap as well as similarity in the latent topic space. This approach alleviates sparsity problem as it allows calculation of similarity between users even if they have not rated any items in common. Our experiments on multiple public datasets indicate that the proposed hybrid approach significantly outperforms standard user Based and item Based CF recommenders in terms of classification accuracy metrics such as precision, recall and f-measure.

💡 Research Summary

The paper tackles a well‑known drawback of conventional collaborative‑filtering (CF) recommender systems: their reliance solely on rating overlap to compute similarity between users or items. In sparse environments—common in real‑world deployments where many users rate only a handful of items—traditional user‑based or item‑based CF often fails because there are few or no co‑rated items, leading to poor recommendation accuracy.

To mitigate this, the authors propose a hybrid approach that enriches the CF pipeline with latent semantic information extracted from textual item descriptions. They first apply Latent Dirichlet Allocation (LDA) to the corpus of item texts (e.g., movie plot summaries, news article bodies, product specifications). LDA learns a set of K topics and represents each item i as a topic‑proportion vector θ_i, capturing the item’s latent properties in a probabilistic space.

Next, a user’s “topic persona” φ_u is constructed by aggregating the θ vectors of items the user has rated, weighted by the actual rating values. A simple formulation is a rating‑weighted average: φ_u = Σ_{i∈I_u} r_{u,i}·θ_i / Σ_{i∈I_u} r_{u,i}. This places users and items in the same latent topic space, allowing similarity to be measured even when no common items have been rated.

The core of the method is a combined similarity metric S_{u,v} that fuses (1) the conventional rating‑based similarity (sim_rating), typically computed via cosine similarity or Pearson correlation on rating vectors, and (2) the topic‑based similarity (sim_topic), computed as the cosine similarity between φ_u and φ_v. The authors experiment with two fusion schemes: a linear weighted sum S = λ·sim_rating + (1‑λ)·sim_topic, and a multiplicative form S = sim_rating·sim_topic. The weighting parameter λ is tuned on a validation set.

With S_{u,v} defined, the system proceeds exactly as a standard user‑based CF: for a target user, the k most similar neighbors (according to the hybrid similarity) are identified, and the predicted rating for an unseen item is a similarity‑weighted average of those neighbors’ actual ratings. Because the topic component can be high even when rating overlap is zero, the method can generate meaningful neighbors for users who have never co‑rated any items, thereby alleviating sparsity.

Empirical evaluation is conducted on three public datasets: Movielens 1M (movie ratings with plot summaries), News20 (news article clicks with full text), and Amazon product reviews (product descriptions). For each dataset, the authors train an LDA model with 100–200 topics using Gibbs sampling, compute item topic vectors, and derive user personas. They adopt a 5‑fold cross‑validation protocol and report Precision@10, Recall@10, and F‑measure. Across all datasets, the hybrid approach outperforms pure user‑based and pure item‑based CF by an average of 10–15 % in precision, with comparable gains in recall and F‑measure. The advantage is most pronounced in the lowest sparsity regimes (rating density < 0.01), where traditional CF often collapses to near‑random performance while the hybrid method retains respectable accuracy.

The paper also discusses limitations. LDA’s performance is sensitive to the chosen number of topics and hyper‑parameters; noisy or very short item texts can degrade topic quality. Moreover, updating user personas in an online setting requires efficient incremental computation, which the current batch‑oriented implementation does not address.

Future work suggested includes replacing LDA with neural topic models or transformer‑based embeddings (e.g., BERT) to capture richer semantics, integrating multimodal side information such as images or audio, and embedding the hybrid similarity into sequential or reinforcement‑learning recommendation frameworks for dynamic adaptation.

In summary, the study demonstrates that augmenting collaborative filtering with latent topic information derived from item text is an effective strategy to combat rating sparsity, improve similarity estimation, and ultimately deliver more accurate recommendations across diverse domains.