MIM: Multimodal Content-Interest Modeling¶

Core mechanism evidence. Paper and local evaluations are separate.

Paper information¶

Field	Value
Paper	arXiv 2502.00321
Affiliations	Alibaba / Taobao
First public release	2025-02-01 (arXiv v1)
Authors' code	Authors' release.
Adapter / method	`mim`
Local code	`src/auto_research/reproductions/mim/`

Original paper¶

Background and contribution¶

ID-only models struggle with fresh content. Reconstruct masked multimodal input, align content and collaborative interests with SFT, and fuse content/ID interests with CiUBM.

flowchart LR
  C[Content] --> M[Masked reconstruction] --> A[Content and collaboration alignment] --> F[CiUBM fusion]
  I[ID interest] --> F

Key figure from the paper¶

Original Figure 2: MIM. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism¶

\[ \mathcal L=\mathcal L_{\mathrm{mask}}+\lambda\mathcal L_{\mathrm{align}}+ \mathcal L_{\mathrm{rank}},\qquad s(u,i)=g(i_u^{ID},i_u^{content})^\top e_i . \]

Paper results¶

Private Taobao offline MIM/CiUBM beat ID/content baselines; tables do not support one pooled lift. Live CTR +14.14%, RPM +4.12%.

Local reproduction¶

MovieLens-1M, 420 users/640 items, leave-two-out, full catalog. Against ID co-occurrence: Hit@10 0.14048→0.12143, NDCG@10 0.07514→0.07584 (+0.94%), MRR@10 0.05541→0.06228. Metrics.

auto-research reproduce --paper mim --dataset-dir data --seed 42

Reproduction limits¶

Genres proxy image/text. No private impressions, actual multimodal encoder, or serving. Do not compare directly with live effects.