A Practical Guide to Recommendation System Architecture and Methods

1. Overview

This article offers a brief introduction to recommendation systems, explaining their purpose, basic framework, and key components.

2. Product Recommendation System

2.1 Definition of Recommendation System

Recommendation systems address information overload by helping users discover items of interest and uncover latent preferences.

2.2 Architecture

The core pipeline consists of three stages: recall, ranking, and re‑ranking.

Request Flow

When a user opens a page, the front‑end sends user identifiers (e.g., PIN or UUID) to the back‑end. The back‑end selects strategy configurations via AB testing, then invokes the appropriate recall, ranking, and re‑ranking modules. Multiple recall paths generate candidate items, which are sorted, enriched with price and image information, and finally presented. User interactions are logged for later analysis.

Why Use a Recall‑Ranking‑Re‑ranking Funnel?

From a performance perspective, it reduces the candidate set from millions to a few items, controlling the inference cost of complex ranking models. From a goal perspective, each module has distinct responsibilities.

Recall

Recall quickly filters a large item pool to a manageable candidate set. Multi‑path recall combines different strategies or simple models to improve recall rate while maintaining speed.

2.3.1 Advantages and Disadvantages of Multi‑Path Recall

High recall and speed, but each path requires manual tuning and may produce redundant candidates.

2.3.2 Recall Categories

Recall can be divided into non‑personalized and personalized approaches.

Non‑personalized recall: hot items, new arrivals.

Personalized recall: tag‑based, region‑based, collaborative filtering.

Collaborative Filtering

CF includes user‑based, item‑based, and model‑based methods that compute item similarity using cosine similarity or Jaccard index.

Jaccard formula: J(A,B) = |A∩B| / |A∪B|.

Vector Recall

Vectorized recall learns low‑dimensional embeddings for users and items, turning recall into a nearest‑neighbor search, which improves generalization and diversity.

Training can use models such as word2vec; online retrieval employs ANN algorithms like LSH, HNSW, or product quantization.

2.4 Ranking

Ranking refines candidates, with coarse ranking handling large volumes and fine ranking delivering the final order. Wide & Deep models combine a linear “wide” part for memorization with a deep neural part for generalization.

2.5 Re‑ranking

Re‑ranking adjusts the fine‑ranked list to achieve global optimality and meet business goals, using algorithms such as Maximal Marginal Relevance (MMR) to balance relevance and diversity.

def MMR(itemScoreDict, similarityMatrix, lambdaConstant=0.5, topN=20):
    s, r = [], list(itemScoreDict.keys())
    while len(r) > 0:
        score = 0
        selectOne = None
        for i in r:
            firstPart = itemScoreDict[i]
            secondPart = 0
            for j in s:
                sim2 = similarityMatrix[i][j]
                if sim2 > secondPart:
                    secondPart = sim2
            equationScore = lambdaConstant * (firstPart - (1 - lambdaConstant) * secondPart)
            if equationScore > score:
                score = equationScore
                selectOne = i
        if selectOne is None:
            selectOne = i
        r.remove(selectOne)
        s.append(selectOne)
    return (s, s[:topN])[topN > len(s)]

3. Conclusion

The article provides a high‑level overview of recommendation systems, outlining their architecture and key modules, and encourages further study and refinement in practical applications.