Learning from Sequences: A Review of Machine Learning and Deep Learning for Text, Speech, Behavioural, Financial and Genomic Data

Abstract

Much of the data that matter in computing arrive as sequences: the words a user types, the sounds of a language, the purchases in a basket, the daily prices of a stock and the bases of a genome. In each case, the meaning of an element depends on the elements around it. Methods for such data have moved from frequency counts and association rules to recurrent networks, attention and transformers, yet studies in language technology, behavioural analytics, finance and bioinformatics rarely refer to each other. This paper reviews learning methods for sequential and set-structured data in four areas: text and speech, user behaviour and transactions, financial time series and credit, and genomic and molecular sequences for medicine and drug discovery. It places each study on a common path from representation to model to decision. The review finds that association rules and tree ensembles remain useful for transaction and credit data, that recurrent networks such as LSTMs are a strong baseline for price movement, that reinforcement learning helps when a suggestion changes what the user does next, and that attention-based models now lead on long text, speech synthesis and drug-target interaction. Common problems include limited labelled data for low-resource languages and rare diseases, distribution shift over time and the difficulty of explaining sequence models. Future work should share representations across domains, evaluate models on data from later periods and report explanations next to accuracy.