Fig 1.
An illustrative diagram shows how to use MIRN in an industrial recommendation system.
MIRN focuses on the matching stage, which aims to retrieve user-interested items efficiently. Note that the matching stage is concerned with retrieving good items, rather than specific item ranks.
Table 1.
Glossary.
Fig 2.
MIRN takes a sequence of user profile features and user behaviors as input and outputs multiple user representation vectors for item retrieval. User behaviors and other user features (gender, age, etc.) are fed to the multi-interest representation learning module through the corresponding operations (feature embedding and average pooling) of the embedding module. Then, the interest Capsules generated by the multi-interest representation learning module are stitched with the user’s other feature embeddings separately, and multiple user representation vectors are obtained through several Full connection & ReLU layers. Finally, multiple user representation vectors are used in parallel for Top N item retrieval. The Capsule-aware module is mainly used to guide the training of the model.
Table 2.
Statistics of the two datasets for offline evaluation.
Table 3.
Performance comparison of different models on publicly available datasets.
Table 4.
Performance comparison of different models on other Amazon datasets.
Fig 3.
Comparison of HR@50 performance between our model and other models on different datasets.
Table 5.
Model performance for different parameters .
Fig 4.
Performance comparison of different numbers of interest in MIRN.
The y-axis represents the value of HR@50 and the x-axis gives the number of iterations. MIRN performs better with bigger .
Fig 5.
Heatmap of coupling coefficients for user behaviors.
Each behavior has a corresponding coupling coefficient in the corresponding capsule of interest.
Table 6.
Comparison of the speed and accuracy of different dense retrieval models on Amazon Electron.