How KOBE Transforms Personalized Recommendation Reason Generation with Transformers

1. Introduction

In e‑commerce, simple product recommendation is no longer enough; content‑based recommendation requires high‑quality reasons to help consumers decide, spark interest, and increase diversity. The goal in the Spring Festival cloud theme is to generate knowledge‑based personalized recommendation reasons.

Challenges include (1) producing grammatically correct and fluent text, (2) ensuring the reason aligns with product attributes and domain knowledge, and (3) personalizing the reason for different user groups.

Traditional text generation relied on RNN‑based Seq2Seq+Attention. The Transformer (Vaswani et al., 2017) achieves better speed and performance, so we adopt it as the baseline and propose KOBE (KnOwledge‑Based pErsonalized) to incorporate knowledge and personalization.

2. Problem Definition

Given a product name x (character‑level tokenized, length n ), the system should generate a recommendation reason y (length m ) related to the product’s category. The training objective is to make y as close as possible to the reference answer.

We extend the input with feature attributes a (aspects such as appearance, quality, and user categories) and external knowledge w from CN‑DBpedia. The model must produce a personalized, knowledge‑enriched reason y .

3. KOBE Model Design

3.1 Transformer

The Transformer encoder‑decoder uses multi‑layer self‑attention (typically six layers) with positional encoding, followed by feed‑forward networks. The decoder adds cross‑attention to the encoder outputs.

3.2 Attribute Fusion

We embed aspect and user‑category features separately, average them to obtain an attribute vector, and add this vector to each token embedding of the product title before feeding into the Transformer. This enables diverse, user‑specific text generation.

3.3 Knowledge Incorporation

For each token of the product title we retrieve textual descriptions from CN‑DBpedia, sample five concepts per entity, and concatenate them as knowledge input. A separate knowledge encoder (self‑attention) produces a representation u , which is combined with the title representation h via bidirectional attention (title‑to‑knowledge and knowledge‑to‑title) to obtain a fused representation.

4. Spring Festival Cloud Theme Deployment

During the Spring Festival, KOBE was deployed to generate personalized reasons for each category, producing fluent, knowledge‑aware, and engaging texts.

5. Results & Outlook

Recommendation‑reason generation is maturing; content‑driven e‑commerce will benefit from such techniques, and further exploration can aid platforms.

Reference: Dzmitry Bahdanau et al., 2014; Thang Luong et al., 2015; Minjoon Seo et al., 2016; Ilya Sutskever et al., 2014; Ashish Vaswani et al., 2017; Bo Xu et al., 2017.