• The ClustPTF remarkably improves recommendation diversity and predictive performance. • Sentiment analysis is used to reduce model sparsity by complementing missing ratings. • Two data structures of multi-criteria preferences are introduced for clustering them. • Experimental results show ClustPTF’s potential on a near real-time recommendation. In the recommender system field, diversity as the measure of recommendation quality has gained much attention recently. However, many pieces of research have shown that it has a trade-off relation with predictive performance. To improve recommendation diversity and predictive performance in multi-criteria recommender systems, we propose a clustering-based parallel tensor factorization (ClustPTF). In the ClustPTF, sentiment analysis alleviates model sparsity, and the K-means clustering considering rating behaviors groups similar user preferences into sub-models and leads to improve recommendation diversity. The sub-models are then factorized in parallel to predict ratings in near real-time. With one dataset gathered from TripAdvisor, experiments showed that the ClustPTF considerably improve recommendation diversity (13.44x of a conventional tensor factorization (TF 0 )) and response time (23.13x of the TF 0 ). Even its predictive performance is superior to the TF 0 (41.06% improvement in MAE). Furthermore, the ClustPTF outperformed recent techniques in recommendation diversity and predictive performance (i.e., MAE and precision).