Empowering Time Series AI: How Salesforce is taking advantage of synthetic data to improve base models

Time series analysis faces significant obstacles in data availability, quality and diversity, critical factors in the development of effective base models. Real world data sets often fall short due to regulatory limitations, inherent biases, poor quality and limited related textual annotations, which makes it difficult to create robust and generalizable time series models (TSFM) and time series models based on large language models (TSLLM). This scarcity affects tasks such as prognosis, classification, anomalies detection, reasoning and subtitle, which limits the complete potential of current advances in artificial intelligence.

The ai ​​of Salesforce has addressed these challenges by proposing a comprehensive approach to take advantage of synthetic data to improve TSFM and TSLLM. His recent study, “Empowering the analysis of time series with synthetic data”, presents a novel strategy to use synthetic data to improve models training, evaluation and adjustment, focusing on mitigating biases, increasing the diversity of data sets and enriching contextual information. In developing innovative data generation frames and incorporating synthetic data sets, Salesforce ai aims to advance the practical application of TSFM and TSLLM, especially in confidential domains such as medical care and finance, where data exchange is very regulated.

The technical cornerstone of the Salesforce ai Research methodology involves several synthetic data generation approaches, each that addresses specific aspects of temporal series dynamics, such as trends, seasonal patterns and noise characteristics. For example, the ForecastPFN method combines exponential linear and periodic seasonal trends with noise distributed by Weibull, effectively simulating realistic but diverse scenarios. Similarly, Timesfm integrates linear trends by parts and models of mobile average (weapon) self -regressive with periodic patterns. Another innovative technique, Kernelsynth of Chronos, uses Gaussian processes (GPS) combined with linear, newspaper and radial -based nuclei (RBF) to generate rich synthetic data sets. These methods allow a creation of controlled but varied synthetic data that helps capture an integral range of realistic time series behaviors.

The findings of the Salesforce team highlight substantial benefits derived from synthetic data in multiple stages of model development. In the prestrénmente, synthetic data sets provided clear performance improvements, notably demonstrated in models such as Frevastpfn, Mamba4cast and Timesfm. For example, the previous forecast previously in synthetic data showed significant improvements in zero shooting forecast scenarios, while the Cronos found optimal performance gains when mixing about 10% of synthetic data with real world data sets, beyond which additional synthetic data could potentially degrade performance due to less diverse representations. In addition, synthetic data also played a crucial role in the evaluation, allowing researchers to precisely evaluate the model's capacities, understand internal representations and identify gaps in the patterns learned. Moment used synthetic sinusoidal waves to evaluate internal incrustations and model sensitivity to variations in the characteristics of the temporal series, which demonstrates its effectiveness in the capture of subtle tendencies and frequencies.

The document also addresses current limitations in the use of synthetic data, identifying areas for future improvement. A critical gap is the absence of systematic integration methods for synthetic data sets, which suggests the need for structured frames to identify and fill the data patterns of the real world strategically missing. Another limitation observed is the mastery of statistical methods, which causes a call to explore generative techniques based on data, such as diffusion models, to improve realism. Salesforce researchers further emphasize the non -exploit potential in the use of synthetic data during the fine adjustment phases to address specific domain gaps or model weaknesses in a more efficient and adaptive way.

In conclusion, Salesforce ai Research shows that synthetic data offers a powerful set of tools to overcome data -related challenges in time series analysis. By systematically integrating high quality synthetic data sets in several stages of model development, TSFM and TSLLM can achieve improved generalization, reduced biases and improved performance in various analytical tasks. Despite the existing limitations, such as guaranteeing realism and alignment, proactive advance and exploration of synthetic data generation methodologies indicate significant potential. Future research, as Salesforce suggests, should focus on improving data realism, systematically addressing data gaps and exploiting iterative human synthetic data generation processes in the circuit. These advances could drastically expand the applicability and reliability of time series models, establishing a solid basis for future innovations in artificial intelligence.

Verify he Paper. All credit for this investigation goes to the researchers of this project. In addition, feel free to follow us <a target="_blank" href="https://x.com/intent/follow?screen_name=marktechpost” target=”_blank” rel=”noreferrer noopener”>twitter And don't forget to join our 85k+ ml of submen.