CONVERSION FORECASTING IN A DIGITAL ENVIRONMENT BASED ON AN MLP+LSTM MODEL WITH INTEGRATION OF STABILITY AND EVOLUTION INDICATORS OF BEHAVIORAL SEGMENTS

Abstract

The article proposes an approach to conversion forecasting in a digital environment based on a hybrid MLP+LSTM model with the integration of behavioral segment stability and evolution indicators. The relevance of the study is determined by the growing need to improve predictive analytics in digital services through the joint consideration of aggregated session characteristics, the temporal logic of user interaction, and the structural properties of behavioral segments to which user sessions belong. Unlike conventional approaches that rely only on tabular session-level features or only on sequential behavioral patterns, the proposed framework combines both types of information and supplements them with segment-dynamic indicators derived from cluster stability and cluster evolution analysis.

At the first stage of the study, user sessions are segmented by means of cluster analysis in order to identify relatively homogeneous behavioral groups within the digital environment. Such segmentation makes it possible to move from isolated session-level analysis to the examination of stable and unstable behavioral subsystems that reflect recurring interaction patterns. At the second stage, the structural stability of the identified segments is assessed using the CSAI index, which provides a quantitative basis for evaluating the internal consistency and resilience of each behavioral cluster. At the third stage, the MONIC model is applied to analyze the evolution of segments across temporal slices of the digital environment. This allows the detection of segment preservation, transition tendencies, merging, splitting, and other forms of structural transformation that may affect conversion behavior.

The indicators obtained from CSAI and MONIC are not treated only as auxiliary analytical results, but are directly integrated into the predictive contour in the form of additional segment-dynamic features. These features extend the descriptive space of the model and allow it to account not only for the current characteristics of an individual session, but also for the stability of the segment to which the session belongs and the potential direction of its structural transformation. In this way, the model incorporates both micro-level behavioral evidence and meso-level segment dynamics.

The predictive module is implemented as a hybrid MLP+LSTM architecture. The MLP component processes aggregated session-level features together with segment-dynamic parameters generated from CSAI and MONIC. The LSTM component processes the real user clickstream as an ordered sequence of actions performed within a session. The fusion of these two branches enables the model to simultaneously capture the integral session profile, the temporal structure of user interaction, and the stability and transformation properties of the corresponding behavioral segment. As a result, the proposed approach creates a more informative basis for forecasting conversion-related outcomes in complex digital systems.

The practical significance of the study lies in the possibility of improving conversion prediction accuracy, identifying unstable or transformation-prone behavioral segments, and supporting decision-making related to personalization, interface optimization, user journey redesign, and adaptive management of digital services. The proposed approach may also serve as a methodological foundation for broader decision support systems focused on behavioral analytics in complex digital environments.