Multi-Criteria Recommender Systems (MCRS) play an important role in delivering more personalized and accurate recommendations by considering multiple aspects of user prefer ences rather than a single overall rating. However, MCRS often suffers from high data sparsity, which hampers the accuracy of rating predictions. To address this, many studies have employed techniques like Singular Value Decomposition (SVD) and Au toencoder (AE), each with its limitations: SVD tends to produce inaccurate predictions when preference patterns are complex and non-linear, while AE risks failing to learn latent representations optimally on highly sparse data. This study proposes a sequential hybrid model that integrates AE and SVD, where AE first reconstructs the rating matrix by modeling non-linear patterns, followed by SVD to refine predictions using dominant linear structure. A case study is conducted using two multi-criteria datasets, BeerAdvocate and OpenTable, each with five evaluation criteria per item. Experiments include tuning activation functions and optimizers for the Autoencoder. The proposed AE-SVD model consistently outperforms standalone AE, standalone SVD, and the SVD-AE variant. On BeerAdvocate dataset, AE-SVD achieves an MAE of 0.4763, RMSE of 0.6325, Precision@10 of 0.3225, and Recall@10 of 0.6341. On OpenTable dataset, it attains an MAE of 0.6868, RMSE of 0.8813, Precision@10 of 0.1409, and Recall@10 of 0.8478. Additionally, on a cold-start subset of the BeerAdvocate dataset (users with limited rating history), the model maintains reasonable accuracy, with an MAE of 0.5657 and RMSE of 0.7510. These findings indicate that combining AE and SVD sequentially improves prediction accuracy in sparse multi-criteria recommender systems.