Abstract
Recommendation systems have become essential tools for navigating modern digital life. Traditional methods often focus on historical user-item interactions to determine past preferences, while sequential recommendation systems analyze recent interaction sequences to capture short-term interests. However, these sequential methods often overlook the functional relationships between item properties. This paper proposes the CERTU (Convolutional Embedding Recommendation with Time and User Preference) model, which integrates time characteristics and user preferences. By analyzing diverse relationships between items, the model captures dynamic preferences over time and considers the influence of both individual and multiple items within a sequence. Experimental results demonstrate that the CERTU model significantly outperforms current baseline methods.
Introduction
Traditional recommendation systems frequently struggle with two extremes: focusing solely on static, long-term preferences or over-emphasizing the most recent clicks. This often leads to a failure in modeling time-varying item relationships. For example, a user's interest in complementary items (like a phone case after buying a smartphone) behaves differently than their interest in substitutive items (like another smartphone). CERTU addresses this by bridging the gap between long-term user personality and short-term functional dynamics.
Methodology
The CERTU model architecture consists of three primary layers designed to process item sequences and knowledge graph data.
- Data Input & Knowledge Graph Layer: The model uses a Knowledge Graph to categorize item relationships. It models complementary interest using a normal distribution (peaking shortly after purchase) and substitutive interest using a dual normal distribution (accounting for the immediate drop and eventual rise in replacement interest).
- Convolutional Calculation Layer: Treating the sequence of $L$ item embeddings as an $L \times d$ matrix, the model applies Horizontal Convolution to capture joint behaviors and Vertical Convolution to capture the influence of individual items.
- Full Connection Layer: This layer fuses the short-term sequential features from the CNN with a long-term user representation vector to produce the final recommendation.
Experimental Setup
The model was evaluated using the Amazon Cellphones and Accessories dataset, comprising 227,900 users, 10,300 items, and 193,200 entries. The researchers utilized buy_after_viewing and also_buy data to define the substitution and complementation relationships essential for the model's temporal functions.
Results and Analysis
CERTU was compared against seven baselines, including BPR, Caser, and GRU4Rec. It achieved superior performance with a Hit Ratio (HR@10) of 0.5497 and an NDCG@10 of 0.3524.
A critical part of the evaluation was the ablation study, which tested the model's components. The authors compared the full CERTU model against CERTU-I (ignoring item relationships) and CERTU-T (ignoring time).
As shown in the results, the full CERTU model outperforms both variants, proving that the synergy between temporal modeling and item relationship awareness is the key driver of its success.
Conclusion
The CERTU model demonstrates the power of combining convolutional neural networks with domain-specific knowledge about item relationships and temporal dynamics. While the current model uses normal distributions to approximate interest, future work could explore more complex, non-linear interest curves and deeper graph-based relationship mining to further refine recommendation accuracy.
Summary updated