Why Deep Learning Is Revolutionizing Recommendation Systems

Why introduce deep learning into recommendation systems?

Deep learning offers strong representation capabilities, robustness to noisy data, the ability to model sequential information with RNNs, more accurate user and item feature learning, and unified data processing, making it a powerful tool for modern recommender systems.

Directly extract features from content with strong representation power

Robust to noisy data

RNNs can model dynamic or sequential data

Learn user and item features more accurately

Facilitate unified processing of heterogeneous data

Deep Recommendation Systems

Deep learning has rapidly advanced in NLP, image processing, and reinforcement learning, and is now being applied to recommendation systems across five major categories, with four highlighted here.

Learning item embeddings

Deep collaborative filtering

Feature extraction directly from content

Session‑based recommendation with RNNs

Hybrid combination algorithms

1. Learning item embeddings & 2VEC models

Embedding learns a new vector representation from input data, allowing the original vectors to be expressed in a more compact space.

1.1 Embedding as MF

Matrix Factorization (MF) learns latent user and item vectors; embeddings serve as these latent features, enabling item‑to‑item recommendation.

Learning similar features

Deep models can learn user, item, and preference vectors analogous to MF, treating MF as a simple neural network where one‑hot user IDs are inputs and hidden layers represent latent features.

1.2 Word2Vec

Word2Vec encodes words into vectors (e.g., "example" → [0.44, 0.11]) by learning semantic relationships from large corpora, using Skip‑Gram or Continuous Bag‑of‑Words (CBOW) models.

Word2Vec‑CBOW

CBOW predicts a target word from its surrounding context, converting one‑hot vectors into embeddings, then compressing context information before applying softmax for probability estimation.

1.3 xxx2vec

Beyond basic Word2Vec, many variants (Paragraph2Vec, Content2Vec, Meta‑Prod2Vec) incorporate additional context such as paragraphs, product metadata, or higher‑dimensional items, adapting to specific recommendation scenarios like short videos versus movies.

2. Deep Collaborative Filtering (DCL)

2.1 Auto‑encoders

Auto‑encoders (AE) are unsupervised models that reconstruct their input, learning latent user or item vectors. They predict missing ratings by minimizing reconstruction loss on the rating matrix.

Stacked Denoising Auto‑encoders (SDAE) add noise to inputs, forcing the encoder to learn robust representations; Bayesian SDAE further models user and item latent vectors with Gaussian priors.

Collaborative Recurrent Auto‑encoder (CRAE)

CRAE replaces shallow encoders with recurrent neural networks (e.g., GRU, LSTM) to capture sequential user behavior, improving recommendation for session data.

2.2 DeepCF

DeepCF models (e.g., MV‑DNN, TDSSM, Co‑evolving Features, Product Neural Network) combine deep neural networks with collaborative filtering to model user‑item interactions, temporal dynamics, and product hierarchy.

2.3 Google Recommendations

Google’s Wide & Deep model merges linear and deep components for feature interaction; YouTube’s recommender uses DNNs to embed user watch history, search queries, and side information, producing top‑N video recommendations.

Content Features in Recommenders

Deep learning can extract high‑dimensional features from images, text, and audio, which are then combined with collaborative filtering (CF) and content‑based filtering (CBF) for hybrid recommendation.

Session‑based Neural Recommendation

Session‑based recommendation treats a user’s interaction sequence as a time‑ordered session. Traditional content‑based and collaborative filtering methods ignore item order, whereas RNNs (GRU/LSTM) capture sequential preferences and predict click‑through rates.

Summary

This article reviews various deep learning applications in recommendation systems, showing that models such as DNNs, auto‑encoders, and CNNs can be effectively employed across different domains, and emphasizing the growing importance of deep learning as a core technology for future recommender systems.

Reference

O. Barkan, N. Koenigstein: ITEM2VEC: Neural item embedding for collaborative filtering. IEEE 26th International Workshop on Machine Learning for Signal Processing (MLSP 2016).

M. Grbovic et al.: E‑commerce in Your Inbox: Product Recommendations at Scale. 21st ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD’15).

Q. Le, T. Mikolov: Distributed Representations of Sentences and Documents. 31st International Conference on Machine Learning (ICML 2014).

T. Mikolov et al.: Efficient Estimation of Word Representations in Vector Space. ICLR 2013 Workshop.

T. Mikolov et al.: Distributed Representations of Words and Phrases and Their Compositionality. 26th Advances in Neural Information Processing Systems (NIPS 2013).

F. Morin, Y. Bengio: Hierarchical probabilistic neural network language model. International workshop on artificial intelligence and statistics, 2005.

F. Vasile et al.: Meta‑Prod2Vec – Product Embeddings Using Side‑Information for Recommendations. 10th ACM Conference on Recommender Systems (RecSys’16).

Source: https://zhuanlan.zhihu.com/p/33214451