Personalized Recommendation and Advertising Algorithms for E‑commerce: Business Overview, Recall and Ranking Optimization, Multi‑Task Modeling, and Future Directions

The presentation, delivered by Li Xinru, an algorithm engineer at JD.com, introduces the role of personalized recommendation in e‑commerce, emphasizing how recommendation algorithms directly affect user experience, retention, and platform revenue.

Business Overview : The system serves multiple pages such as installment mall recommendation, floor recommendation, JD Outstanding, and marketing recommendation. Key metrics include CTR, GMV, ARPU, order count, add‑to‑cart, favorites, dwell time, and shares. The recommendation pipeline extracts user profiles, invokes several recall services, aggregates results, and passes them to a ranking service for feature engineering and online prediction.

Recall Optimization :

Profile recall is enhanced by a pre‑recall + model approach, replacing rule‑based scoring with a learned model (LR → FFM → NN). Negative samples include both exposure data and recall results, improving coverage.

Similarity recall leverages item‑based collaborative filtering, embedding methods (MF, ALS, word2vec, tag2vec, random‑walk, node2vec) and evaluates embeddings via online i2i recall, offline loss/AUC, and precision/recall/F1.

To reduce online load, a unified similarity path is built by training a model that predicts similarity scores offline, then serving a single recall stream.

Ranking Optimization :

Sample re‑weighting addresses class imbalance by assigning higher weights to conversion‑related actions and applying log‑smooth scaling.

Loss weighting (focus loss) amplifies errors on hard samples, improving CTR and multi‑objective performance.

Multi‑model ensembles include independent models combined post‑prediction (CTR × CVR for order‑oriented goals, price‑aware adjustments for GMV‑oriented goals) and hard‑share architectures similar to ESMM, with additional tricks such as fixing CTR parameters, using Swish activation, and soft‑share (MMOE + attention) to share knowledge across tasks.

Recommendation + Advertising Integration : The same recommendation infrastructure supports advertising scenarios, including RTA (real‑time advertising) and DSP. Techniques such as CPM‑based bidding, CPC control, and ROI‑driven pricing reuse multi‑task models.

Cold‑Start for Advertising : To mitigate scarce conversion samples, exposure data from partial competition samples and multi‑task modeling are used. Data augmentation methods (SMOTE, GAN, WGAN) generate synthetic samples, with discriminators jointly predicting authenticity and conversion probability.

CPC Control : A PID controller adjusts bid prices to meet target CPC, balancing proportional, integral, and derivative terms to reduce static error, overshoot, and oscillation.

Conversion Delay Modeling : An exponential decay model predicts conversion probability over days, using a combined loss for binary conversion and time‑to‑conversion, improving long‑tail attribution.

Results & Future Plans : Multi‑objective ranking boosted CTR, UV value, and GMV; DSP saw higher click rates, lower CPC, and improved ROI. RTA conversion rates increased up to 15%. Future work includes uplift modeling, conversion‑delay refinement, multi‑task ROI estimation, automated hyper‑parameter search (CEM, genetic algorithms), and industry‑specific feature mining.