<p>Video recommendation has become an important way to assist people in exploring and identifying enormous videos they are mostly interested. Conventional systems typically rely on user-video interactions and specific content attributes to generate recommendations. However, when these attributes are unavailable or incomplete, the performance of classical models tends to worsen significantly. Moreover, a growing challenge arises from the larger proportion of unregistered users on video platforms, which further complicates accurate recommendations. To address these challenges, we introduce the convolutional Kronecker forward Harmonic network (CKFHNet) for video recommendation. Initially, the input video usage file is acquired from the KuaiRand dataset. After that, the user and video accumulation matrix is generated. Then, user grouping and video groupings are performed using Bayesian fuzzy clustering. Next, the CKFHNet model is trained based on user group data. Thereafter, query matching is achieved through Mahalanobis distance to identify relevant user groups. Mahalanobis distance provides a more precise identification of relevant user groups based on their feature distributions. Afterwards, the recommendation of relevant videos is performed by employing CKFHNet. The CKFHNet is designed by incorporating a convolutional neural network (CNN) and deep Kronecker network (DKN) with forward harmonic analysis. Finally, the ranking of recommended videos is executed. CNNs efficiently extract visual features from the video content, DKN improves the model's ability to capture complex relationships between users and videos, and forward harmonic analysis assists in modelling temporal patterns within the data. The devised CKFHNet model obtained a maximal accuracy of about 92.204% as well as a minimal normalized mean squared error of 0.065 and a normalized root mean squared error of 0.256. This result demonstrates that the devised CKFHNet model not only outperforms traditional models but also provides highly reliable and precise video recommendations. The performance improvement obtained by the proposed CKFHNet model in terms of accuracy while comparing with the existing methods graph neural network, multilayer feed-forward network, multi-feature video recommendation method based on hypergraph convolution, NetRec framework, deep spiking parallel convolutional neural network, graph neural networks multi-modal recommendation, deconfounding causal recommendation-mixture-of-experts, and dynamic micro-video recommendation framework are given by 13.66%, 11.93%, 9.44%, 12.51%, 9.20%, 7.00%, 3.90%, 2.24% respectively.</p>