Improving Advertising Inventory Forecasting with Deep Spatial‑Temporal Tensor Factorization

Every time a user visits a web page they encounter ads, and both advertisers and ad platforms need to estimate future ad exposure, known as "ad inventory forecasting," to set pricing and plan campaigns.

This task is difficult because user traffic varies over time (daily, weekly, seasonal) and because modern ad delivery relies on detailed user segmentation and tagging, which adds complexity to the prediction problem.

For example, estimating how many users will see a shampoo ad next month requires not only historical volume data but also an understanding of which user groups the new campaign will target.

In the machine‑learning field, ad inventory forecasting is essentially a time‑series prediction problem: using historical data to forecast future values. It also must incorporate user‑category features such as location, gender, and age.

Prediction methods have evolved from traditional models like ARIMA, GARCH, and TBATS to deep neural networks based on LSTM and CNN, which capture patterns more effectively.

Despite progress, current models still face three major challenges: (1) handling the massive scale of data and attribute dimensions, (2) dealing with complex and sometimes sparse attribute tag combinations, and (3) capturing both short‑term and long‑term seasonal cycles.

To address these issues, researchers from Tencent and the University of Science and Technology of China proposed a new model described in the paper "Large‑scale User Visits Understanding and Forecasting with Deep Spatial‑Temporal Tensor Factorization Framework." The model uses a deep spatio‑temporal tensor factorization architecture, an attention‑based embedding for user attributes, and a multi‑task training scheme, enabling it to disentangle attribute effects, short‑term trends, and long‑term cycles.

Experiments on a three‑year, billions‑of‑visits dataset (with over 10,000 distinct attribute combinations) showed that the model reduced prediction standard deviation by 15.6% versus ARIMA, 8.7% versus CNN‑based deep networks, and 5.8% versus matrix‑factorization‑plus‑deep‑learning baselines. It also remained robust when 20% of training data were missing and required only about one‑third of the parameters of comparable CNN models, saving computational resources.

More accurate inventory forecasts enable advertisers to target audiences more precisely, improve campaign planning, and increase ad order fulfillment rates, which in turn boosts platform revenue and advertiser satisfaction.

The model has already been deployed in Tencent’s online advertising system, delivering noticeable revenue growth, and the authors will present a detailed technical walkthrough in an upcoming WizTalk live session.