SRS: A recommendation system for scientific publication

A core step in any research project is identifying prior work that provides a reliable view of the state of the art. This activity can be partially automated using artificial intelligence, particularly recommender systems to streamline literature discovery. This paper introduces SRS (Scientific Recommendation System), a recommendation system for scientific publications that integrates scholarly metadata through documented, open sources and supports exploratory searches in topics such as resilience, automated planning, cloud computing, HTN planning, and vision-language-model planning. SRS implements a two-stage hybrid approach: (i) a content-based candidate generation stage that computes semantic similarity between the user query and paper text metadata (title, abstract, and keywords) using TF–IDF vectorization and cosine similarity, and (ii) a lightweight bibliometric re-ranking stage based on publication year and citation count (min–max normalized) to prioritize temporally and impact-adjacent papers. The interface provides visual analytics that compare yearly distributions of recommended papers against query-driven results and reports cosine-similarity scores and histogram frequencies of similarity/proximity values for transparency. Validation is conducted under a two-level protocol. Core effectiveness results are reported on a fully labeled five-query subset (E1–E5) using rank-aware metrics (P@10, P@20, NDCG@10, MAP@20, and MRR) computed offline from a frozen snapshot (snapshot v1, 17 Mar 2026). In addition, an expanded protocol of 25 queries spanning AI Planning and AI Health is documented to strengthen external coverage, including a 10-query overlapped subset with graded relevance (0/1/2) for inter-annotator agreement analysis using Cohen's κ and graded NDCG reporting. In the main evaluation, query-specific recovered pools are denoted by | D q |, whereas N is reserved exclusively for the scalability experiment. NDCG@10 ranges from 0.5833 to 1.0000 and Recall@ | D q | ranges from 0.0006 to 0.5000, indicating that SRS can retrieve a high proportion of relevant publications while reducing manual effort in early-stage literature exploration. Scalability and sensitivity tests are also reported. These findings suggest that transparent, source-flexible recommendation pipelines can support early-stage literature mapping; broader validation across domains and user populations remains future work. Licensed commercial indexes (e.g., Scopus/Web of Science) are supported as optional connectors but are not required for the reproducible evaluation.