Understanding Novel Literature Recommendation: Characteristics, Tagging Challenges, and Multi‑Modal AI Algorithms

The presentation begins by defining novels as a literary genre that combines character development, background description, and narrative storytelling, highlighting their large volume, diverse characters, and often lengthy chapters, which pose specific challenges for recommendation systems.

From a recommendation perspective, novels exhibit three main challenges: massive content size, high user decision cost due to long reading cycles, and the prevalence of low‑quality, homogeneous works that make quality discrimination essential.

Tag‑based representation, while common, suffers from semantic redundancy (e.g., overlapping tags like "romance" and "love"), difficulty in computing tag weights, and the inability to quantify similarity between works, especially since most users only deeply read one or two novels.

The article then outlines the reader’s decision process: users first encounter a list of novels with cover, title, rating, summary, and tags, decide to click for details, and finally read the content, emphasizing the need for recommendation algorithms to guide users smoothly through these stages.

To model novels effectively, multi‑modal feature representations are introduced, combining textual, visual, and behavioral signals. Simple concatenation of vectors leads to redundancy, so advanced multi‑modal fusion methods are employed to reduce noise and improve accuracy.

A dual‑tower (user‑item) architecture is described, where the left tower encodes user information and the right tower encodes novel content, with convolutional and fully‑connected layers producing semantic vectors for similarity matching.

After training, semantic clustering visualizations demonstrate that novels with similar meanings are grouped together, confirming the effectiveness of the learned representations.

For recall modeling, a YouTube‑style deep neural network (DNN) is applied, consisting of two layers (representation and fully‑connected) and a final softmax layer that extracts user semantic vectors for efficient candidate retrieval.

The talk concludes with a thank‑you note and a call for audience engagement.