Designing Deep Learning Models for Item Similarity in Recommendation Systems

Introduction

Item‑based collaborative filtering was first proposed by Amazon in 2001 and remains widely used because of its speed, simplicity and interpretability. Traditional similarity is computed with Pearson correlation, which does not optimise any objective function.

Problem

How can we design a deep‑learning model that computes item similarity from user behaviour data?

Analysis and Answer

We can approach the problem from unsupervised and supervised perspectives.

Unsupervised learning

The common pipeline is to embed items into low‑dimensional dense vectors and measure similarity with cosine distance. Techniques from NLP and network embedding such as prod2vec, item2vec and Graph Convolutional Neural Networks are applicable. Figure 1 illustrates a skip‑gram based prod2vec model.

Supervised learning

Supervised methods integrate the similarity model with the downstream recommendation task. The model takes two item IDs as input and outputs a similarity score; during inference the scores for all items a user has interacted with are summed to obtain a final ranking, as shown in Figure 2.

Loss design and training

Collect positive and negative feedback and use the predicted scores as logits for a loss function. With a sigmoid output this corresponds to binary cross‑entropy; alternatively, negative sampling with multi‑class cross‑entropy or learning‑to‑rank losses can be employed. After training, the model can accept any pair of items and output their similarity.

References

[1] Sarwar et al., “Item‑based collaborative filtering recommendation algorithms,” WWW, 2001.

[2] Grbovic et al., “E‑commerce in your inbox: Product recommendations at scale,” KDD, 2015.

[3] Barkan & Koenigstein, “Item2vec: neural item embedding for collaborative filtering,” MLSP, 2016.

[4] Monti et al., “Geometric matrix completion with recurrent multi‑graph neural networks,” NeurIPS, 2017.

[5] Ying et al., “Graph convolutional neural networks for web‑scale recommender systems,” KDD, 2018.