Didi’s “Guess Where You’re Going” Feature: Product Benefits and Bayesian Gaussian Modeling

The article introduces Didi’s “Guess Where You’re Going” functionality, which proactively suggests a destination (POI) to users before they type anything, aiming to reduce input effort and enhance user experience.

From a product perspective, the feature offers two main advantages: it shortens the user’s interaction by allowing a direct click on a correctly guessed POI, and it creates a memorable, personalized impression that can increase user loyalty.

Technically, the problem is framed as a highly personalized prediction task where the user already knows the intended destination; therefore, the algorithm focuses on accurately recalling that POI rather than recommending items based on broader interests.

The design relies on two key observations: (1) the candidate POIs for recall are the user’s frequently visited locations, and (2) the most influential contextual factors are departure time, departure location, and whether the day is a workday.

Didi adopts a straightforward Bayesian approach: for each user, a Gaussian distribution is fitted to model the conditional probability of a destination given the context. The model computes P(Y = y_i | X) by estimating the prior frequency of each POI and the likelihood of the context X (time, location, weekday) under a Gaussian assumption.

Estimating the Gaussian parameters requires handling the 24‑hour periodic nature of time. The mean μ is obtained by minimizing the sum of squared circular distances (ensuring differences never exceed 12 hours), while the variance is derived from the dispersion of observed departure times around μ.

At inference time, the system evaluates the probability of each candidate POI under the current context, ranks them, and selects the top‑N destinations that exceed a predefined confidence threshold.

The authors acknowledge several limitations: the model uses only a few contextual features, ignoring richer signals such as weather; it does not incorporate user behavior trajectories beyond the immediate context; and extending the approach would require more comprehensive data and raise privacy concerns.

Reference: Zhang L, Hu T, Min Y, et al. A Taxi Order Dispatch Model based On Combinatorial Optimization. ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2017.