Telecom providers rely on innumerable data points to optimize network performance, predict demand, analyze usage patterns, monitor signal coverage, ensure reliability and detect fraud, among other factors.

But this spatial and time series data — based on geographic boundaries and events occurring over time, respectively — is complex and can be difficult for nontechnical users to analyze.

“Analyzing these two types of data requires specialized skills due to their complexity and interdisciplinary nature,” Phil Darringer, Kinetica VP of product management at real-time database platform Kinetica, told SDxCentral.

In hopes to bridge this gap, Kinetica today announced the availability of Kinetica SQL-GPT for telecom.

The platform consists of a database engine specifically designed for time-series and spatial analysis, and a user interface connected to a large language model (LLM) converts plain English into Structured Query Language (SQL).

“This combination of these two capabilities enable telco professionals to have a conversation with their data,” said Darringer.

Telecoms use both spatial data and time-series data to predict demand for better resource allocation, analyze usage patterns, determine service requests and experiences, monitor signal coverage for targeted infrastructure improvements and detect fraud through location-based anomalies, Darringer said.

Spatial data identifies longitude and latitude or boundaries of objects or phenomena. It involves intricate geographic relationships that require an understanding of spatial concepts including distances, proximities and topologies.

Meanwhile, time series data involves observations or measurements collected over successive time intervals. It demands expertise in handling temporal dependencies that impact forecasting, trend interpretation and seasonality adjustments.

Understanding and optimally using both types of data often requires an interdisciplinary approach based on user knowledge in areas including statistics, geography and computer science as well as expertise in fields such as urban planning, environmental science, finance and network infrastructure.

“Combining spatial and time series data requires skills to manage and integrate diverse datasets, accounting for both spatial and temporal dimensions,” said Darringer. This includes the ability to integrate diverse datasets and knowledge in creating maps, graphs and visual representations.

Call logs, for example, record attributes such as signal strength, which changes from one moment to the next. Spatial elements involve physical transmission and signal reception as a phone moves, and tower locations create a spatial network that influences signal strength and call quality.

Consider, for instance, a mobile user makes a series of calls throughout the day, Darringer said. Each call (which is time-series) occurs in different geographies (which is spatial). Analyzing this data ultimately reveals temporal patterns linked to geographies, “emphasizing how telecom data inherently intertwines time and space,” he said.

Traditional analytic databases weren’t primarily built to efficiently handle spatial and time-series data, said Darringer.

They lack capabilities for complex “joins,” such as longitude and latitude in one table to a polygon in another. They also cannot match records in one dataset to the closest preceding or succeeding record in another based on specific timestamps or intervals.

Additionally, existing infrastructure has “rudimentary or no specialized functions” for spatial and temporal analytics. This includes, for instance, identifying proximity or shortest path between the two objects, aggregating within spatial boundaries or calculating line of sight between towers.

Legacy databases require “scarce skills” to manually code intricate queries, and they can only answer basic queries (such as ‘how many customers do we have’) where the answer is known in advance, Darringer noted. Also, the underlying data is stale, day’s or week’s old.

“As a result, much of the value from the data goes to waste as the volume of questions that the organization would like to ask is severely constrained,” said Darringer.

Kinetica’s real-time GPU database was purpose-built for spatial and time-series workloads, according to the company. With its specific telecommunications platform, users can ask complex questions and visualize and interact with billions of data points on a map.

Compared to traditional engines that spit out simplistic, outdated information, SQL-GPT for Telco reflects data as it is generated from the source, Darringer said.

For instance, field technicians might ask “What five pieces of equipment in my area are causing the highest number of dropped calls?’ and immediately receive a prioritized list of those requiring attention.

They can then ask follow up questions such as, “Only show me pieces of equipment where we have replacement parts in local inventory,” or “sort the equipment requiring attention to reduce total miles driven.”

Or, a user in an office can analyze millions of device connections to pinpoint abrupt drops in download speeds in a particular area — say, Nashville, which experienced a 30% dip on Oct 13 and Oct 26. They can then identify specific buildings that were impacted and prioritize the 20 highrises where ground crews should be dispatched.

T-Mobile, for its part, processed 90 billion spatial object records with Kinetica, turning latitudes and longitudes into routes to be used for coverage planning and ROI projections for network and retail build out.

While previously it would have taken several weeks and “significant technical churning” to map just two streets in San Francisco, Kinetica mapped all streets in the entire state of California in roughly an hour.

“Our analysis must be predictive driven,” James No, T-Mobile software engineer, commented in a case study. “It’s no longer sufficient to only use a static historical perspective to inform network planning.”

Jake Holman, senior manager for software development at T-Mobile, further explained that with Kinetica,  “today, we’re blending more big data sets leading to better performing models and more accurate ROI prediction.”

Another inherent challenge, according to Darringer, is that the telecom industry uses specialized language, terms and jargon — such as quality of service (QoS) around network performance based on bandwidth, latency and reliability.

Kinetica’s LLM was fine-tuned to this industry-specific vernacular, is available in a containerized, secure environment either on-premises or in the cloud and runs natively — along with the database engine — on Nvidia GPUs.

Darringer said that, “this gives the platform a level of performance and ability to handle complex and novel questions that aren’t possible using traditional CPU based approaches.”