Intent Scoring

Intent scoring operates as a low-latency inference pipeline, recalculating scores on every new behavioral event. Unlike batch segmentation, the model ingests streaming clickstream data and updates a user's intent probability within milliseconds. This requires a feature serving layer that combines pre-computed user embeddings with real-time session features—such as dwell time, scroll velocity, and add-to-cart events—to generate a calibrated probability score between 0 and 1.

The score is not a static snapshot; it is a function of the ordered sequence of actions leading to the current moment. Architectures like Behavior Sequence Transformers (BST) or Long Short-Term Memory (LSTM) networks process the temporal trajectory of clicks, views, and searches. This captures concept drift in a session—for example, distinguishing a user who is casually browsing from one whose behavior pattern is accelerating toward a purchase intent.

A single intent score often masks multiple underlying objectives. Production systems decompose intent into distinct, calibrated probabilities:

• Purchase Intent: Likelihood of completing a transaction within the session.
• Churn Intent: Probability of abandoning the platform in the next N minutes.
• Category Affinity: Strength of preference for a specific product taxonomy. Each sub-score is independently calibrated using Platt scaling or isotonic regression to ensure the output is a true probability.

For anonymous or first-visit users with no historical data, intent scoring relies on in-session behavioral velocity and contextual priors. Features such as the rate of page views, referral source, device type, and geographic location provide an initial signal. The system rapidly transitions from a population-level prior to a personalized posterior as the user generates more intra-session events, often within 3-5 clicks.

The intent score is not an end in itself; it is a decisioning signal fed into a real-time orchestration engine. When a user's score crosses a predefined threshold, it triggers automated actions:

• High Intent (>0.8): Trigger a live chat invitation or a limited-time discount.
• Medium Intent (0.4-0.8): Surface a personalized product recommendation carousel.
• Low Intent (<0.4): Serve educational content or a newsletter signup prompt. This threshold-based actuation closes the loop between prediction and business outcome.

User intent is inherently non-stationary. A high purchase intent score from 30 minutes ago is meaningless if the user has since gone idle. Scoring models apply time-decay weighting—often an exponential decay function with a half-life measured in minutes—to historical events. This ensures that the most recent actions, such as a sudden increase in dwell time on a pricing page, dominate the score calculation, while stale signals rapidly lose influence.