Neural IR Meets Graph Embedding: A Ranking Model for Product Search

Abstract

Recently, neural models for information retrieval are becoming increasingly popular. They provide effective approaches for product search due to their competitive advantages in semantic matching. However, it is challenging to use graph-based features, though proved very useful in IR literature, in these neural approaches. In this paper, we leverage the recent advances in graph embedding techniques to enable neural retrieval models to exploit graphstructured data for automatic feature extraction. The proposed approach can not only help to overcome the long-tail problem of click-through data, but also incorporate external heterogeneous information to improve search results. Extensive experiments on a real-world e-commerce dataset demonstrate significant improvement achieved by our proposed approach over multiple strong baselines both as an individual retrieval model and as a feature used in learning-to-rank frameworks. CCS CONCEPTS• Information systems → Retrieval models and ranking; Information retrieval query processing;

Temporal-Contextual Recommendation in Real-Time

“…To study the effects of IPS off-policy correction, we imputed a global impression bias by counting item frequency in the training set. Figure 4(a) shows that the NDCG with our meta-data model is comparable with SOTA from Zhang et al [35] (green versus dotdash blue line at the top). Using cold-start models hurt performance because there is no significant distribution shift between the train and test splits, i.e., total-variation loss is lower between train/test (15%) than two random folds splitting the test set (18%).…”

“…7 To measure meta data performance, we use bag-of-words (BoW) features and a product search dataset. Following Zhang et al [35], we consider 37.7k train queries and 16k test queries as users, and all 184k unique products as candidates. Query/product features come from their bag-of-words (BoW) representation, where we used only the top 3000 frequent words and performed a square-root transform on the BoW features.…”

Ranking Associative Entities in Knowledge Graph by Graphical Modeling of Frequent Patterns

“…Graph embedding techniques like DeepWalk [23] are effective for association analysis in graphical structures [3], in which low-dimensional representations of the nodes with neighboring and cooccurrence relations are learned. Zhang et al [29] propose a graph embeddingbased neural ranking framework to overcome the query-entity sparsity problem by integrating features in click-graph data. On heterogeneous information networks, recent studies for proximity search [18] learn graph embedding models to rank associative nodes by given semantic relations.…”

Distilling Structured Knowledge into Embeddings for Explainable and Accurate Recommendation

“…As discussed in [46], the mutual learning framework can help to find more robust local minimal by entropy regularization. Different from DML which jointly trains two neural architectures (eg., Resnet and MobileNet), our approach combines two types of models, i.e., embedding/neural models and path/graph models, with rather different inductive biases [45] to allow them to benefit more from each other.…”