<p>Nonnegative CP decomposition (NCPD) is an important method for dimensionality reduction for nonnegative multi-way data, which has a wide range of applications in machine learning, signal processing, image processing, and pattern mining. One well-known algorithm for NCPD is the hierarchical alternating least squares (HALS). However, it is challenged when dealing with large-scale data because of poor convergence and significant cost. To address this issue, we introduce a set of learnable semi-orthogonal transforms to compress the large-scale tensor into a small-scale one, and propose the learnable transform-based NCPD, which is also incorporated the column unit constraints to avoid numerical ill-condition and enhance computational stability. The approximation errors due to the semi-orthogonal transforms are discussed, and the corresponding HALS, or rather a proximal alternating minimization-based algorithm is presented to compute the decomposition. Furthermore, we also employ multiple inner iterations and the extrapolation strategy to further improve the aforementioned algorithm. The convergence and complexity analyses of the proposed algorithms are provided. Extensive numerical experiments on synthetic and real data, and the applications in clustering are conducted to illustrate the superior performance of our decomposition and algorithms.</p>