This paper, titled "Intent-Enhanced Data Augmentation for Sequential Recommendation," presents a new framework called IESRec. It aims to move beyond the "randomness" of traditional data augmentation in recommendation systems by injecting semantically meaningful "intent segments" into user behavior sequences.
TL;DR
IESRec improves sequential recommendation by replacing random data augmentation (like random masking or cropping) with intent-enhanced insertion. By strategically inserting "plausible" artificial sequences into user histories, the model creates high-quality positive and negative samples that help a Transformer-based encoder better capture dynamic user interests through joint contrastive learning.
What Problem Does This Paper Address?
Sequential recommendation systems try to predict a user's next move (e.g., what item they will click next) based on their chronological history. Because user data is often sparse, researchers use Data Augmentation (DA) to create more training examples.
However, most current DA methods rely on randomness: they randomly delete, reorder, or mask items in a sequence. The authors argue this has two major flaws:
- Noise Injection: Randomly deleting a key item might destroy the very signal that defines a user's intent, leading to "blurred" representations.
- Underutilization: Many models only use augmented data for auxiliary tasks (like contrastive learning) rather than directly training the main recommendation task on them.
Methods
The core innovation of IESRec is Intent Insertion. Instead of random changes, the model generates "artificial" sequences based on real item transitions found in the dataset.
- Intent Sequence Generation: For any given item, the model looks at its "adjacency set"—the items that frequently follow it in the training data. It then "walks" through these sets to build a chain of items that represents a plausible user path.
- Positive Sample Augmentation: The model inserts these artificial chains into non-terminal positions in a user's history. This reinforces the underlying intent without changing the final "target" item the model is trying to predict. These samples are mixed directly into the training set.
- Negative Sample Augmentation: The model appends an artificial chain to the end of the sequence. Since this "future" path is artificial and doesn't lead to the actual ground-truth next item, it serves as a "hard negative" sample.
- Joint Training: The model uses a Transformer encoder and a dual-loss strategy. It minimizes a standard recommendation loss (Cross-Entropy) on both original and positive-augmented data, while simultaneously minimizing a Contrastive Loss that pulls original sequences closer to their positive versions and pushes them away from the negative ones.
Datasets
The authors evaluated IESRec on three popular subsets of the Amazon dataset:
- Beauty: Personal care products (relatively sparse).
- Clothing: Apparel and accessories (highly sensitive to time/seasonality).
- Sports: Sporting goods and outdoor equipment.
All datasets were preprocessed to remove "cold-start" users and items (those with fewer than five interactions), and sequences were capped at a maximum length of 50.
Key Insights and Ideas
- Plausibility over Randomness: By using "neighbor search" to build augmented sequences, the model ensures that the new data "looks" like real user behavior. This reduces the noise that typically plagues random augmentation.
- Hard Negatives are Key: In contrastive learning, distinguishing between a real sequence and a "plausible-looking but wrong" sequence (the intent-inserted negative) forces the model to learn much finer-grained features of user intent.
- Direct Supervision: Unlike many contrastive models that treat augmented data as a side-task, IESRec treats positive augmented sequences as valid training data for the main recommendation task, effectively "densifying" the sparse interaction matrix.
Key Results
IESRec consistently outperformed state-of-the-art baselines (including MAERec, DCRec, and CL4SRec) across all three datasets.
- Accuracy Gains: On the Beauty dataset, IESRec achieved a Hit Rate (HR@50) of 0.1818, significantly higher than the best-performing baseline, MAERec (0.1787).
- Ranking Quality: The model showed strong improvements in NDCG (Normalized Discounted Cumulative Gain), suggesting it isn't just finding the right items, but ranking them higher in the list.
- Ablation Success: The authors proved that both the positive samples (PS) and the contrastive learning (CL) module are essential. Removing the CL module caused a "sharp decline" in accuracy, particularly on the Sports dataset.
Limitations and Future Work
- Fixed Hyperparameters: The paper uses a fixed length ($K$) for the inserted intent segments. In reality, user intent spans vary; some intents are short-lived (buying a gift), while others are long-term (a hobby). Future work could explore dynamic segment lengths.
- Computational Overhead: Generating these intent sequences requires traversing adjacency sets and performing neighbor searches. While effective, the paper does not deeply discuss the scalability of this generation process for datasets with millions of items.
- Semantic Depth: The "intent" is currently defined by item-to-item transition frequency. Incorporating actual item metadata (like categories or text descriptions) could make the intent segments even more accurate.
Summary updated