Modeling Stage-wise Evolution of User Interests for News Recommendation

📝 Paper Summary

News Recommendation Temporal Graph Learning

SEIN improves news recommendation by partitioning user history into temporal stages to jointly model stable global preferences via graph learning and evolving short-term interests via LSTM and self-attention.

Core Problem

Existing news recommenders typically rely on either static global graphs (missing short-term evolution) or pure sequential models (missing long-term stability), failing to capture how reading habits adapt to emerging events.

Why it matters:

News is highly time-sensitive; user interests shift rapidly due to breaking events (e.g., pandemics, elections), rendering static profiles outdated.
Purely sequential models often over-emphasize recent clicks, forgetting stable long-term preferences (e.g., a consistent interest in finance despite a temporary sports spike).
Balancing stability and plasticity is critical for accuracy, yet prior methods struggle to integrate global collaborative signals with fine-grained temporal dynamics.

Concrete Example: A user typically reads sports news (stable interest). During a pandemic, they intensely consume health news (short-term shift). A static graph model might keep recommending sports during the crisis, while a sequential model might completely forget the sports interest after the crisis subsides. SEIN handles both by modeling the transition.

Key Novelty

Unified Global-Local Temporal Framework (SEIN)

**Stage-wise Segmentation**: Partitions continuous user interaction history into discrete temporal subgraphs (e.g., weekly windows) to capture coherent local interests rather than treating history as a flat sequence.
**Dual-Perspective Modeling**: Combines a Global Preference Module (LightGCN on full graph) for stable habits with a Local Preference Module that tracks evolution across stages.
**Hybrid Evolution Mechanics**: The local module uses an LSTM branch for progressive, step-by-step interest shifts and a Self-Attention branch to aggregate long-range dependencies across historical stages.

Breakthrough Assessment

7/10

Addresses a logical gap in news recommendation by structurally separating stable and evolving interests. The stage-wise subgraph approach is a sensible evolution of dynamic graph methods, though the components (LightGCN, LSTM) are standard.

⚙️ Technical Details

Problem Definition

Setting: Personalized News Recommendation

Inputs: User set U, News set I, chronological click history for each user segmented into time intervals

Outputs: Probability that user u will click on candidate news item i

Pipeline Flow

Data Preprocessing: Convert click timeline into T discrete time intervals
Global Modeling: LightGCN on full User-News Graph → Global Embeddings
Local Modeling: Construct T temporal subgraphs; Initialize with Global Embeddings
Evolution Learning: Parallel LSTM (progressive) and Self-Attention (long-range) branches
Fusion & Prediction: Combine Global, LSTM, and Attention embeddings → Score

System Modules

Global Preference Modeling (GPM)

Capture long-term, stable collaborative signals from the entire history

Model or implementation: LightGCN

Local Preference Modeling (LPM) - Initialization

Initialize stage-specific learning with global priors

Model or implementation: GCN (per stage)

Short-term Evolution Branch

Model progressive, step-by-step changes in interest

Model or implementation: LSTM

Long-range Aggregation Branch

Capture dependencies across all historical stages (e.g., recurring seasonal interests)

Model or implementation: Self-Attention Layers

Fusion & Prediction

Combine signals and predict click probability

Model or implementation: Concatenation + Dot Product

Novel Architectural Elements

Hybrid temporal architecture: Parallel LSTM and Self-Attention branches operating on GCN-encoded temporal subgraphs
Global-to-Local initialization scheme where global graph embeddings initialize local temporal subgraphs

Modeling

Base Model: LightGCN (Global Encoder)

Training Method: Joint optimization of Recommendation Loss + Regularization

Objective Functions:

Purpose: Optimize recommendation accuracy.

Formally: Binary Cross-Entropy Loss L_rec = - sum (y log y_hat + (1-y) log (1-y_hat))
Purpose: Align local temporal representations with global stable preferences.

Formally: Contrastive Consistency Loss L_cl = - log ( exp(sim(e_local, e_global)/tau) / sum(exp(sim(e_local, e_neg)/tau)) )
Purpose: Ensure smooth evolution of interests between adjacent time steps.

Formally: Smoothness Loss L_sl = sum || e_u^t - e_u^{t-1} ||^2

Key Hyperparameters:

tau: Temperature for contrastive loss (value not reported)
lambda_t: Weight for recommendation loss
lambda_cl: Weight for consistency regularization
+ 1 more
lambda_sl: Weight for smoothness regularization

Compute: Not reported in the paper

Comparison to Prior Work

vs. LightGCN: SEIN adds temporal dynamics via stage-wise subgraphs; LightGCN is static.
vs. LSTUR: SEIN uses graph learning for the long-term component instead of a simple ID embedding, and models stage-wise transitions rather than item-by-item sequences.
vs. TCCM: SEIN explicitly models progressive evolution (LSTM) alongside contrastive alignment, whereas TCCM focuses primarily on causal/contrastive signals.
+ 1 more
vs. DGSR [not cited in paper]: DGSR updates graphs dynamically per interaction; SEIN uses discrete stage-wise subgraphs for coarser, more coherent temporal modeling.

Limitations

Relies on pre-defined temporal intervals (e.g., weeks), which may not align perfectly with actual user interest shifts.
Requires maintaining both a global graph and multiple temporal subgraphs, which likely increases memory overhead.
The smoothness regularization assumes gradual interest changes, which might conflict with sudden breaking news events.

Reproducibility

Code: https://github.com/yikeJin/SEIN

Code is publicly available at https://github.com/yikeJin/SEIN. The paper provides the mathematical formulation clearly, but specific hyperparameter values (learning rate, batch size, embedding dimension) are not included in the provided text snippets.

📊 Experiments & Results

Evaluation Setup

News Recommendation on large-scale real-world datasets

Benchmarks:

Not specifically named in text (News Click Prediction)

Metrics:

Not explicitly reported in the paper
Statistical methodology: Not explicitly reported in the paper

Main Takeaways

The paper claims the unified framework consistently outperforms strong baselines (like DIGAT, NRMG, TCCM) across diverse user behaviors.
Ablation studies (qualitatively described) verify the contribution of both the global preference module and the local temporal evolution branches.
The approach effectively balances long-term stability (via Global Graph) and short-term plasticity (via Local Subgraphs), addressing the 'stability-plasticity' dilemma in news recommendation.
Note: Specific numeric results were not contained in the provided text snippets, so quantitative takeaways are omitted.

📚 Prerequisite Knowledge

Prerequisites

Graph Convolutional Networks (GCN)
Sequence Modeling (LSTM, Attention)
Contrastive Learning

Key Terms

LightGCN: A simplified Graph Convolutional Network that learns user/item embeddings by linearly propagating signals over the interaction graph without heavy non-linearities

LSTM: Long Short-Term Memory—a recurrent neural network architecture designed to model sequential data and capture dependencies over time

Self-Attention: A mechanism that relates different positions of a sequence to compute a representation of the sequence, allowing the model to weigh the importance of past interactions adaptively

Stage-wise Evolution: Modeling user interests not as a continuous stream but as a sequence of distinct phases (stages), capturing how preferences transition from one period to the next

Temporal Subgraph: A subset of the user-item interaction graph containing only interactions that occurred within a specific time window

Consistency Regularization: A training objective that forces representations learned from different views (e.g., global graph vs. local time window) to be similar, ensuring alignment

Smoothness Regularization: A constraint enforcing that user preferences do not change abruptly between adjacent time steps, smoothing the evolutionary trajectory