Deep Learning Based Search Ranking Optimization for 58.com Rental Services

Deep learning’s strong feature representation ability has driven breakthroughs in vision, text, and speech, and it is now a key technology for commercial search ranking. 58.com, the largest Chinese lifestyle information platform, applies deep learning to its rental‑search service to improve user‑item matching, increase click‑through and conversion rates, and boost overall platform revenue.

The typical user flow involves searching for a specific community or subway station, retrieving a set of relevant posts, and ranking them before display. Effective ranking requires modeling diverse features, capturing user interests, and aligning results with user intent.

The ranking system consists of a data layer and a strategy layer. The data layer aggregates logs from front‑ and back‑end, extracts and processes features, and generates training samples via positive‑negative sampling. The strategy layer combines machine‑learning/deep‑learning models with business rules to produce robust ranking scores.

Model evolution follows a path from traditional LR/FM/XGBoost to deep architectures:

Wide&Deep : combines a linear LR component (memorization of low‑level feature co‑occurrences) with a DNN component (generalization of high‑level embeddings).

DeepFM : replaces the LR part with a Factorization Machine to automatically capture second‑order feature interactions while sharing embeddings with the DNN.

DCN : further replaces the LR part with a cross network that explicitly models higher‑order feature crosses.

DIN : introduces an attention mechanism to weight user historical behaviors according to the candidate item, reducing reliance on handcrafted statistics.

DIEN : adds an interest‑evolution layer (GRU + AUGRU) to model the temporal dynamics of user interests, improving both CTR and CVR predictions.

Feature construction includes user‑dimensional, post‑dimensional, and contextual features. User features cover IDs, timestamps, queries, and behavior statistics; post features include IDs, location, price, area, orientation, etc.; contextual features provide real‑time and historical CTR/CVR statistics and exposure metadata.

Offline sample generation follows a pipeline of sampling, feature extraction, feature engineering (normalization, bucketing, hashing), and model training. Positive‑negative sample ratios are kept around 1:4–1:5. Model evaluation uses GAUC, which aggregates per‑user AUCs to better reflect online ranking performance.

Optimization efforts focus on:

Feature handling: one‑hot encode low‑dimensional categorical features, embed high‑dimensional sparse IDs, and concatenate all before the attention layer.

Prediction stability: ensure batch‑normalization, Dice activation, and dropout behave correctly during inference.

Data loading speed: switch to TensorFlow Dataset for parallel preprocessing, raising GPU utilization from 10 % to 50 % and cutting training time from 7 h to 1 h.

Multi‑task learning: jointly train CTR and CVR tasks with shared embeddings and attention, using a weighted sum loss (or ESMM‑style cascade) to improve conversion modeling.

Prediction efficiency: replace List‑based request payloads with primitive arrays, and simplify user‑sequence features by sending only raw data and performing padding on the server side, reducing serialization size by 90 % and cutting latency by ~5 ms.

Online experiments show that the fully optimized multi‑task DIEN model outperforms the baseline XGBoost model in both click‑through and conversion metrics, with stable latency (<12 ms per batch of 20) and timeout rates below 0.2 %.

In conclusion, the end‑to‑end pipeline—from offline data preparation to online inference—demonstrates substantial gains for the rental search scenario. Future work includes further embedding optimization for ultra‑sparse features, incorporating visual CNN features of listings, and extending the approach to other verticals such as recruitment and commercial real estate.

References: Cheng et al., 2016 (Wide&Deep); Guo et al., 2017 (DeepFM); Wang et al., 2017 (DCN); Zhou et al., 2018 (DIN); Zhou et al., 2019 (DIEN); Ma et al., 2018 (ESMM).