← notes

Lifelong User Action Sequence Modeling @Pinterest

2025-08-15 · recsys, sequence-modeling

TransAct V2: Lifelong User Action Sequence Modeling on Pinterest Recommendation (2025)

Ranking Model Architecture

![[transact-v2-ranking-model.png|450]]

LCE=hH{wh[yhlogf(x)h+(1yh)log(1f(x)h)]}L_{CE} = \sum_{h \in H} \{-w_h [y_h \text{log} f(x)_h + (1 - y_h) \text{log}(1 - f(x)_h)]\}

TransAct V2 Architecture

![[transact-v2-architecture.png|475]]

Sall=NN(SLL,c)SRT[:r]NN(SRT[r:],c)NN(Simp,c)S_{\text{all}} = \text{NN}(S_{LL}, c) \oplus S_{RT}[:r] \oplus \text{NN}(S_{RT}[r:], c) \oplus \text{NN}(S_{imp}, c) F=CONCAT(EPinSage(Sall),ec)+Eact+Esurf+EposRS×dF = \text{CONCAT}(E_{PinSage}(S_{all}), e_c) + E_{act} + E_{surf} + E_{pos} \in R^{|S| \times d}

Serving

![[transact-v2-serving.png|525]]

![[transactv2-serving-optimization.png]]

Experiment and Results

![[results-offline.png|450]]

![[results-online.png|350]]