<p>Accurately predicting the distribution of residents’ travel modes is crucial for decision-makers to allocate transportation resources efficiently and improve operational effectiveness. In recent years, deep learning has been successfully employed to automatically extract travel preferences from residents’ behavior. However, existing methods often overlook the temporal characteristics of residents’ travels, leading to models failing to capture the dynamic changes in residents’ preferences. Moreover, the cold start problem, arising from limited travel records, remains a challenge, often leading to a performance drop of over 10% in baseline models. To address these challenges, we propose a model that balances multiple travel factors to more accurately predict travel patterns. To adapt to changes in residents’ preferences, we utilize an LSTM with an attention mechanism (LSTMA) to extract the temporal dependencies in behavior sequences. The cold start problem, characterized by a lack of historical travel records for residents, is addressed by introducing Model-Agnostic Meta-Learning (MAML) to provide personalized preference predictions. To capture the nuanced interplay between resident information and travel behavior, we have developed a Balance Multi-Travel Pattern Prediction Framework (BMTMP) and integrated it with two preceding modules to form the METMP-LSTMA model. The proposed METMP-LSTMA model, grounded in the principles of recommender systems for capturing personal preferences, serves as a foundational technology for future travel recommendation scenarios (e.g., route planning or sustainable mode suggestions). Extensive experiments on two real-world datasets demonstrate that METMP-LSTMA achieves an average accuracy improvement of 5.7% over the average of baselines, with a particularly strong gain of 10.9% for cold-start users.</p>