This study investigates the impact of AutoEncoder-based imputation methods on similarity-based recommendation performance in attribute-driven systems where missing data frequently occurs. Focusing on environments with limited explicit user-item intera...
This study investigates the impact of AutoEncoder-based imputation methods on similarity-based recommendation performance in attribute-driven systems where missing data frequently occurs. Focusing on environments with limited explicit user-item interactions, such as influencer marketing platforms, we empirically compare two missing data strategies: imputing missing values (Imputed) versus preserving missing states (Masked).
We utilized real-world influencer-campaign data (7,811 samples) and the MovieLens 1M dataset, evaluating five AutoEncoder variants: AE, DAE, MAE, VAE, and CVAE. Missing values were artificially introduced at rates of 10%, 30%, and 50% during testing. Imputation accuracy was measured using MAE, RMSE, and Cosine Similarity, while recommendation performance was assessed via Recall@K and NDCG@K. All experiments were repeated 30 times with statistical significance testing.
Results reveal a critical disconnect between imputation accuracy and
recommendation performance. While AutoEncoder models achieved superior numerical reconstruction, this improvement did not consistently translate to better recommendations. Notably, in real-world data, the MAE model—which showed the highest imputation accuracy—yielded the lowest recommendation performance. The Masked approach, which preserves missing values, often matched or exceeded the performance of sophisticated imputation methods. Conversely, MovieLens data showed that imputation improved recommendation quality, demonstrating
dataset-dependent effects.
t-SNE visualization analysis revealed that imputation-induced over-smoothing reduces discriminative power in the embedding space by collapsing user representations toward the mean. In real service data, missing values exhibit Missing Not At Random (MNAR) characteristics that reflect user-specific signals.
Imputing these values removes meaningful information, while the Masked approach preserves uncertainty and maintains structural diversity in the embedding space.
This study demonstrates the non-correlation between imputation accuracy and recommendation performance, highlighting the necessity of differentiated missing data strategies based on the underlying missing mechanism. Our findings provide theoretical and practical insights for designing attribute-based recommendation systems.
Keywords: Recommender Systems, Missing Data, AutoEncoder, Attribute
Embedding, Similarity-based Recommendation, Data Imputation