Using LSTM for Video Recommendation at Tubi: Model, Training, and Results

As a video streaming service, Tubi's core product relies on a recommendation system. To explore whether Long Short‑Term Memory (LSTM) networks, known for success in speech recognition and machine translation, can improve content recommendation, the authors implemented a basic LSTM‑based recommender and omitted Tubi‑specific details for brevity.

The LSTM algorithm, a variant of recurrent neural networks introduced by Hochreiter and Schmidhuber in 1997, addresses the vanishing‑gradient problem and performs well on sequential data such as user activity logs. The experiment uses implicit user feedback—videos watched and watch time—to predict the next most likely video, treating the task as a next‑video prediction while still producing a probability distribution over the entire catalog. User histories are encoded with one‑hot vectors for each video.

The training setup draws inspiration from Andrej Karpathy’s character‑level language model (lecture 10), which predicts the next character from a fixed‑length context. Although video sequences lack the strict ordering constraints of text, the LSTM can still capture user interests from past viewing behavior.

Various history lengths were tested, and the network consists of an LSTM layer followed by dense and softmax layers. Categorical cross‑entropy is used as the loss function and RMSProp as the optimizer.

While cross‑entropy yields high accuracy for predicting the exact next video, this metric does not align perfectly with ranking quality. Offline recommendation metrics such as Mean Average Precision (MAP), Mean Reciprocal Rank (MRR), Precision@N, and NDCG@N (with N = 6, the maximum visible videos on Tubi) showed notable improvements.

Two illustrative cases highlight LSTM strengths: (1) When the catalog lacks many sequential episodes (e.g., the teen comedy series “Fred”), providing the first film in history leads the LSTM to rank “Fred 2” and “Fred 3” as the top recommendations. (2) When a user's history contains diverse genres, the LSTM can shift interests, recommending a modern western‑style series followed by mixed thriller/action titles and finally a pure horror film.

In summary, the LSTM‑based recommendation experiment provided valuable insights and demonstrated the potential of sequence models for video recommendation, encouraging further exploration of advanced machine‑learning techniques in this emerging area.