Combining Privileged Information to Improve Context-Aware Recommender Systems

A recommender system is an information filtering technology which can be used to predict preference ratings of items (products, services, movies, etc) and/or to output a ranking of items that are likely to be of interest to the user. Context-aware recommender systems (CARS) learn and predict the tastes and preferences of users by incorporating available contextual information in the recommendation process. One of the major challenges in context-aware recommender systems research is the lack of automatic methods to obtain contextual information for these systems. Considering this scenario, in this paper, we propose to use contextual information from topic hierarchies of the items (web pages) to improve the performance of context-aware recommender systems. The topic hierarchies are constructed by an extension of the LUPI-based Incremental Hierarchical Clustering method that considers three types of information: traditional bag-of-words (technical information), and the combination of named entities (privileged information I) with domain terms (privileged information II). We evaluated the contextual information in four context-aware recommender systems. Different weights were assigned to each type of information. The empirical results demonstrated that topic hierarchies with the combination of the two kinds of privileged information can provide better recommendations.

💡 Research Summary

The paper addresses a critical gap in context‑aware recommender systems (CARS): the automatic acquisition of item‑level contextual information without relying on explicit user input or pre‑labeled data. The authors propose to extract contextual cues from the textual content of items (web pages) by building topic hierarchies that integrate three sources of information: (1) traditional bag‑of‑words (BoW) representing “technical” information, (2) named entities (NE) as privileged information I, and (3) domain‑specific terms (DT) as privileged information II.

To combine these heterogeneous sources, the authors extend the LUPI‑based Incremental Hierarchical Clustering (LIHC) algorithm, originally designed to merge BoW with a single privileged information stream. In the extended version, the privileged set is split into NE and DT subsets, each producing its own co‑association matrix (M_ne and M_dt). The final consensus matrix is computed as

M_nf = (1‑α)·M_t + β·M_ne + θ·M_dt,

where M_t is the BoW matrix, α controls the overall weight of privileged information, and β, θ (with β+θ=α) allocate importance between NE and DT. Hierarchical clustering on M_nf yields a tree of clusters; the most frequent terms in each cluster become topic labels, which serve as contextual descriptors for the items.

The empirical evaluation uses a Portuguese agribusiness portal dataset comprising 4,659 users, 1,543 web pages, and 15,037 access events. Text preprocessing (stop‑word removal, stemming) and TF‑IDF weighting generate the three representations. Named entities are extracted via a standard NER tool, while domain terms are identified using a domain‑specific terminology extractor.

Four state‑of‑the‑art CARS algorithms—C. Reduction, DaVI‑BEST, Weight‑PoF, and Filter‑PoF—are run with the generated contextual topics. Their performance is compared against a baseline Item‑Based Collaborative Filtering (IBCF) that ignores context. Evaluation metrics include Precision@k and Recall@k across multiple values of α, β, and θ. Results show consistent improvements: when α is set between 0.3 and 0.5 and β≈θ (i.e., NE and DT receive balanced weight), the CARS models achieve 3–7 percentage‑point gains in precision over the baseline. The best results occur when both privileged sources are present, confirming their complementary nature; NE captures proper nouns and temporal expressions, while DT conveys domain‑specific concepts that BoW alone misses. Excessive reliance on privileged information (α>0.7) degrades performance because many documents lack sufficient NE or DT features, leading to sparse co‑association matrices.

Key contributions of the work are:

1. Methodological Extension – Generalizing LIHC to simultaneously incorporate multiple privileged information streams, thereby enriching the semantic representation of items.
2. Parameterization of Information Fusion – Introducing α, β, and θ to flexibly balance technical and privileged cues, enabling adaptation to different domains or data qualities.
3. Empirical Validation – Demonstrating that topic hierarchies built from combined NE and DT improve recommendation accuracy across diverse CARS algorithms.
4. Unsupervised, Label‑Free Approach – Avoiding costly manual labeling or explicit context collection, making the technique scalable to large, heterogeneous corpora.

The paper also acknowledges limitations. The quality of NE and DT extraction directly influences the final recommendation performance; domain‑specific term extraction depends on the completeness of the terminology resource. Moreover, experiments are confined to a single domain (agribusiness), so generalizability to other sectors (e.g., movies, e‑commerce) remains to be proven.

Future research directions include: (a) developing automated methods for building and updating domain term lexicons, (b) learning the fusion weights (α, β, θ) dynamically via optimization or meta‑learning, and (c) integrating user‑level contextual signals (time, location) with the item‑level topics to construct hybrid context models. Such extensions could further boost the effectiveness of context‑aware recommender systems while preserving the unsupervised, scalable nature of the proposed framework.