A Survey of Deep Matching Models for Search and Recommendation

Matching is a fundamental problem in both search and recommendation, measuring the relevance between a document and a query or a user's interest in an item. Machine learning, especially deep learning, has become the state‑of‑the‑art solution for learning matching functions from raw inputs.

The article provides a unified perspective on search and recommendation matching, allowing solutions from both fields to be compared within a single framework. It classifies existing deep learning approaches into two groups: representation‑learning methods and matching‑function‑learning methods.

Beyond search and recommendation, similar matching tasks appear in paraphrase detection, question answering, image captioning, and many other applications, where the goal is to match objects from two different spaces.

Figure 1.1: Unified view of search and recommendation matching.

Input layer receives two matching objects, which may be word embeddings, ID vectors, or feature vectors.

Representation layer transforms input vectors into distributed representations using MLP, CNN, RNN, etc., depending on the data type.

Interaction layer compares the two representations and produces a set of local or global matching signals, often stored as matrices or tensors.

Aggregation layer combines the matching signals into a high‑level matching vector using pooling, concatenation, or other operations.

Output layer takes the high‑level vector and outputs a matching score via a linear model, MLP, neural tensor network, or other neural architecture.

The survey aims to help researchers in the search and recommendation communities gain deeper insight into the space of deep matching models, inspire new ideas, and promote further development of advanced techniques.