A TF-IDF AND LOGISTIC REGRESSION PIPELINE FOR SCHOLARLY ARTICLE CLASSIFICATION AND RECOMMENDATION: IEEE XPLORE BENCHMARK STUDY
Keywords:
Scholarly article classification, scholarly article recommendation, abstract-only text classification, TF-IDF, Logis-tic Regression, IEEE Xplore benchmark, content-based recom-mendation, cosine similarity, calibrated probabilities, classifier-aware re-ranking, supervised machine learning.Abstract
The rapid growth of scholarly publications has made automated topical organization and recommendation essential for efficient literature search. However, most existing approaches treat classification and recommendation as two separate tasks with independent representations. This paper proposes a unified content-based framework in which a single TF-IDF representation of the article abstract drives both multi-class topical classification and top k article recommendation. A new benchmark of 11,744 abstracts is constructed from the IEEE Xplore digital library in six topical queries. The abstract text and the topical query label are retained for every record, so the entire pipeline operates on abstracts alone without titles, author keywords, or indexer-supplied terms. A preliminary confusion analysis reveals that two queries (Big Data Analysis and Cloud Computing) exhibit near-complete vocabulary collapse and are consolidated into a single class, yielding a five-domain benchmark: Big Data & Cloud Computing, Data Science, Robotics, Wireless Communication, and Breast Cancer. On the classification side, five supervised learners (Logistic Regression, Linear SVM, SGD, k-Nearest Neighbours, and Decision Tree) are compared under identical 80/20 stratified hold-out and 10-fold cross-validation protocols. The grid-searched Logistic Regression attains 85.01% accuracy (weighted F1 = 0.850), and a soft-voting ensemble of Logistic Regression, Linear SVM, and SGD reach 85.57% (weighted F1 = 0.855). On the recommendation side, the same TF-IDF representation powers a top-10 recommender that achieves MAP@10 = 0.7664 with pure cosine ranking. Reusing the classi-fier’s calibrated class probabilities to re-rank cosine candidates lifts Precision@10 by +12.5 percentage points (to 0.7715) and MAP@10 by +7.6 points (to 0.8428), with consistent gains on NDCG@10 and MRR. Empirically validating the unified design by showing that classification and recommendation reinforce each other on a shared substrate. The dataset, preprocessing pipeline, trained models, and replication scripts are released to support reproducibility.
