SLi-Rec: Modeling Long and Short-Term User Preferences for Personalized Recommendation

Forum topic · ✨步子哥 · 2025-11-26

Summary

This post reviews SLi-Rec, a recommendation model from Microsoft Research Asia and Shanghai Jiao Tong University that adaptively combines long-term and short-term user preferences. The authors argue traditional sequential models fail on user behavior data due to irregular time intervals and irregular semantics of actions. SLi-Rec introduces two components: TC-SeqRec, a time- and content-aware LSTM that injects time-interval and time-span features into gating, and uses attention over the target item to filter irrelevant history; and an Asymmetric-SVD style long-term model where the user vector is an attention-weighted sum of interacted items. An adaptive fusion controller, conditioned on context such as time gaps and item category, dynamically balances the two preference vectors. On Amazon datasets (Electronics, Movies & TV, CDs & Vinyl) and MSN native ad data, SLi-Rec outperforms 11 baselines including DIEN by 0.8%-2.1% AUC. The article also covers offline/online deployment architecture, negative sampling strategy, and ethical implications.

Key points

Problem: Traditional RNN-based recommenders treat user actions uniformly, ignoring (1) *irregular time intervals* between behaviors and (2) *irregular semantics* — some past actions are relevant to the target item, others are noise.
TC-SeqRec (short-term model): A time- and content-aware LSTM from the paper *Adaptive User Modeling with Long and Short-Term Preferences for Personalized Recommendation* (Yu et al., IJCAI 2019).
Time-aware controller: injects time-interval features δtk = φ(Wδlog(tk − tk−1) + bδ) and time-span features stk = φ(Wslog(tp − tk) + bs) into the cell state and output gate, so recent memories decay gracefully and hot interests propagate faster.
Content-aware controller: attention scores ak = exp(xkWs^x ep) / Σexp(xjWs^x ep) measure semantic relevance between historical items and the target item; attention modulates the cell-state update cek = ak ∗ ck + (1 − ak) ∗ ck−1, amplifying relevant memories and suppressing noise.
The short-term user representation p^short_u = Σ a^s_j hej is a weighted average of all hidden states rather than just the final state.
Long-term model: An Asymmetric-SVD approach — the user has no fixed vector; p^long_u = Σ a^l_j xj is an attention-weighted combination of interacted items. Experiments show fewer than 2% of items receive normalized weights above 0.5, i.e., a small set of "soul items" defines long-term identity.
Adaptive fusion: A controller α = σ(Wm[p^short_u, p^long_u, x_context] + bm) balances the two representations based on context (e.g., time gap, target category). When action intervals are under 1 hour, short-term preference dominates (α near 1); beyond 24 hours, long-term weight rises.

Experimental results

Datasets: three Amazon subsets (Electronics, Movies & TV, CDs & Vinyl), the full Amazon set, and real data from the MSN homepage native ad system (1.92M users / 630K items up to 1M users / 1.2M items).
Metrics: AUC and F1-score for CTR prediction.
SLi-Rec outperformed 11 baseline models on all five datasets, with AUC gains of 0.8%–2.1%. On Electronics, SLi-Rec reached AUC 0.8282 vs. 0.8212 for the next-best T-LSTM; it led DIEN by 3.78% (Electronics) and 3.31% (Movies & TV).
Ablations (Table 4): time-only variant T-SeqRec already improves AUC (0.8709 on Movies); applying attention at the state level beats input-level; averaged hidden states beat the final state.
Fusion study (Table 5): fixed α = 0 (long-term only) performs worst; fixed α = 1 is better; adaptive α is best, adding roughly 0.5%–1.2% over fixed settings.

Industrial deployment

Offline/online split: long-term preferences are computed offline over a longer horizon; short-term preferences are inferred online in real time — analogous to batch settlement vs. ATM withdrawals.
Negative sampling with positive-to-negative ratios around 1:5 (4–9 negatives per positive); 7-day training / 3-day test windows on MSN data capture recent trends.

Limitations and discussion

The model relies on LSTM; the authors did not explore Transformer-based architectures or multimodal inputs.
The post closes with ethical reflections: adaptive fusion allows switching between "conservative" and "exploratory" recommendation, raising questions about filter bubbles and user autonomy.

Core references

1. Yu, Z., Lian, J., Mahmoody, A., Liu, G., & Xie, X. (2019). Adaptive User Modeling with Long and Short-Term Preferences for Personalized Recommendation. *IJCAI*, 4213-4219. 2. Hochreiter, S., & Schmidhuber, J. (1997). Long Short-Term Memory. *Neural Computation*, 9(8), 1735-1780. 3. Koren, Y. (2008). Factorization Meets the Neighborhood: A Multifaceted Collaborative Filtering Model. *KDD*, 426-434. 4. Zhou, G., et al. (2019). Deep Interest Evolution Network for Click-Through Rate Prediction. *AAAI*, 33(01), 5941-5948. 5. Beutel, A., et al. (2018). Latent Cross: Making Use of Context in Recurrent Recommender Systems. *WSDM*, 46-54.