Algorithmic Insights into Free Novel Recommendation: Characteristics, Tagging Challenges, and Multi‑Modal Modeling

Guest and Editor Speaker: Wu Zhenhua, Senior Algorithm Director at Lianshang Literature; Editor: Ma Xiaobao; Platforms: DataFunTalk, AI Enlightener.

Introduction The article shares how Lianshang Literature approaches its free novel business from a recommendation‑system algorithm perspective, covering novel characteristics, algorithmic outcomes, and challenges.

1. What Is a Novel? A novel is a literary genre that reflects life and expresses ideas through characters, settings, and plot, often featuring rich imagination and emotional depth (source: Baidu Encyclopedia).

2. Characteristics of Novel Literature

Large volume with many characters, rich plots, and often serialized chapters.

Readers need a long cognitive process to understand novels, leading to high selection cost.

High homogeneity and quality variance; high‑quality novels have huge value.

These traits pose significant challenges for recommendation systems.

3. Novel Tagging Novels are initially described by tags, but tags suffer from redundancy, imprecise semantics, difficulty in weight calculation, and poor quantification, making similarity measurement and recommendation unreliable.

4. Reading Process and Decision Flow Readers first encounter a list view (cover, title, rating, summary, tags), decide to click, view details, and finally read. Effective recommendation should help users quickly form an accurate understanding and proceed to deep reading.

5. Multi‑Modal Feature Representation Recommendation models combine various signals (click‑through rate, reading rate, depth) into a unified representation, as illustrated in the accompanying diagram.

6. Multi‑Modal Feature Fusion Early approaches concatenated vectors, leading to redundancy. Modern fusion methods reduce redundancy and improve accuracy, as shown in the fusion diagram.

7. End‑to‑End Semantic Modeling: Dual‑Tower A dual‑tower architecture models user and item (novel) information separately, with lower layers for representation and upper layers for convolution, extracting the final user semantic vector from the right tower’s fully‑connected layer.

8. Book Semantic Clustering After training, clustering visualizations reveal semantic similarity among books.

9. Recall Modeling: YouTube DNN User modeling uses a YouTube‑style DNN with two layers (representation and fully‑connected), extracting the final user semantic vector from the softmax‑activated FC layer.

In summary, the presentation outlines the challenges of novel recommendation, proposes multi‑modal representations, dual‑tower semantic modeling, clustering, and DNN‑based recall to improve recommendation accuracy for free novel platforms.

Speaker Bio Wu Zhenhua leads recommendation and search at Lianshang Literature, previously held senior algorithm roles at iQIYI, Inveno, and Intel China, focusing on large‑scale content profiling and transfer learning in recommendation systems.

Thank you for reading. Please like, share, and give a triple‑click support.