Today's big data and AI frameworks face problems like questionable accuracy, shallow data processing depth, black-box in-explainability, and oftentimes low processing speed. This paper summarizes the work of Ultipa, introducing Graph XAI (Graph-augmented AI) and highlighting ADEV (Accuracy, Depth, Explainability, and Velocity). In contrast to many systems that sample data due to inability to traverse datasets thoroughly and quickly, particularly hindered by hotspot supernodes, Ultipa's graph system is designed from data structure and system architecture perspective to allow for ultra-low latency deep penetration, and accuracy is achieved with exhaustive traversal, which also allows for exponentially faster velocity. As graph data are ideally queried and processed using graph query languages and algorithms instead of the two-dimensional SQL and stored procedures, the intuitiveness and explainability are crucial in ensuring ADEV being fulfilled, this paper highlights how Ultipa's graph-native query language facilitates real-time recursive queries like path-finding, K-hopping, auto-networking, or identification of topological structures and communities works hand-in-hand with Ultipa's WebGL-powered graph manager to ensure end-toend celerity and explainability.
1. Introduction: XAI and ADEV XAI originally stands for eXplainable Artificial Intelligence, it signifies the needs for explainability against the results generated by AI, as well as the whitebox explainability of the processes leading to the results. As we are broadening the adoption of AI across all industries, the meaning of XAI transcends explainability, more meanings are added to it, including but not limited to:
Accuracy: the computed results should be adequately accurate.
Depth: the ability to traverse connected data set deeply.
The accuracy problem of AI and big-data systems is
commonly attributed to human ignorance or procedural
unfairness of systems designs [
The depth factor is about how deeply and
thoroughly a system can traverse (or penetrate) the
given dataset. A major weakness with relational
databases and big-data frameworks (being the
foundation of today’s AI systems) is their poor ability to
handle recursive queries, such as joining of tables, due
to effect of cartesian product (cardinality), the
performance degradation is exponential as the number
of tables (roughly equivalent to the depth of the query)
[9] and [
Velocity concerns the speed at which data is
generated, captured, schematized, and processed. The
maximum speed of a system is only tested by putting
into the context of the maximum depth it reaches within
a bounded timeframe. Given the rise of cryptocurrency,
blockchain and Web 3.0, numerous performance
analysis [16] and [17] and [18] have been conducted
against systems implementing such infrastructures.
Most existing big-data and NoSQL frameworks (and the
RDBMS) are designed with storage-centric mindset
where horizontally scalability design is prioritized
higher than depth-oriented velocity. Such mindset
ensures unsatisfactory velocity when there is need for
deep data processing – on the other hand, the popularity
of AI introduces added layers of black-box
inexplainability when processing data through machine
learning or deep learning frameworks [19] and [
This paper focuses on four aspects of XAI which
collectively referred to as ADEV (Accuracy, Depth,
Explainability and Velocity), and how ADEV can be
achieved with a novel graph system design [21].
2. XAI and ADEV How-to
Graph [
Taking K-hop as an example, BFS is guaranteed to be accurate and effective, because DFS does not track the depth of shortest-path (hops) from current vertex to the starting vertex and standing at the Kth hop on a DFS path does not guarantee the correctness. The point here is that it’s rudimentary to open up the implementation core of any analytical query and make it white-box explainable, to ensure the results as well as the procedures are explainable.
The advances of LLM and GPT have given us the impression (or illusion) that AI soon will be taking over average human beings both on IQ and EQ fronts. Scholars around the world have investigated and criticized the hallucination and black-box problems with LLM/GPT [25] and [26] and [32]. From XAI’s perspective, LLM/GPT’s hallucination and unexplainability are rooted in their incapability to conduct deep traversals, or causality searches. Figure 2 illustrates that GPT lacks the ability to conduct causality search, which if conducted in a graph database is to find the shortest path between the parties against the dataset that must be part of what GPT [27] has been trained upon. Figure 3 shows that a shortest-path query of up to 5 hops is conducted on a data set populated with Wikipedia data. • or no de-duplication of K-hop, or partial traversal, all of which may cause the results to be redundant, incomplete, or outright wrong. Ill Product-Market Fit: system designed for batch-processing, pre-calculation and precaching cannot server real-time scenarios.
We use a concrete example to illustrate the accuracy problem when running K-hop queries against a large Twitter-2010 SNS dataset of 42MM vertices and 1470MM edges, the dataset is densely populated with average degree of ~70 (SNS friends of a user), and with the existence of many hotspot supernodes with over 1MM neighbors (high-impact social influencer).
Inadequate computing power: sampling and incomplete traversal of data. Many big-data and AI systems are comfortable handling metadata operations but not networked analytics.
Faulty design or implementation: This happens with unvalidated implementations, such as using DFS for K-hop or shortest-path finding In Table 1, the Twitter dataset’s topology is changed in real-time by connecting a vertex with one of its 3rdhop connected vertices with a new edge (orangecolored, as illustrated in Figure 4) , and we check the starting vertex’s 1st, 3rd and 6th hop neighbors immediately before and after the topology change – in some cases the k-hop results may change dramatically, as shown in the last column in Table 1. In Figure 5, adding an edge between the C001 vertex and C009 will change C001’s 1-hop neighbors from 3 to 4, and 2-hop neighbors unchanged, and 3-hop’s from 7 to 6. If a system uses pre-processing and caching mechanism, it will continue to read pre-stored (stale) results and not be able to churn out updated query results accurately and instantly. At Ultipa, we designed an HTAP system [22] and [30], to ensure changing topologies is accurately and instantly reflected across all system data structures so that graph operations results can be accurate. This is further discussed in the following section.
Data structure plays a pivotal role here, conducting accurate analytics over graphs require both agility and resiliency, such as the ability to handle multi-graph as well as simple-graph, filter by direction of relationships, or attributes tied to vertices and edges, which may affect query results. For instance, if querying K-hop by inbound edges in Figure 4 (with added dashed blue ones), C001 has only one 1-hop neighbor, and one 2-hop neighbor (not the same as K-hop results).
P r e d i c t i o n M a t r i x
Accuracy problem is also tied to complicated system architecture, for instance, accuracy is easier to achieve with single-threading, but once multi-threading and data-partitioning-n-parallel-processing are introduced, data structures and systems architectures are more complex, and the results validation become much harder. Taking Louvain community detection algorithm as an example, the original algorithm was designed to function in a serial fashion, and parallel computing will complicate the matter by yielding faster but possibly inaccurate results (numbers of communities).
Accuracy problem can lead to serious ramifications in real-world applications. Taking retail-bank credit card spending (turnover) prediction as an example, Bigdata/DL frameworks are slow and inaccurate, and each 1% mismatch can be equivalent to $1 billion loss of cash reserve (and in-liquidity). Ultipa models card transactions as a graph network and extracts features via graph queries and algorithms such as weighted node degrees, page-rank, and random-walk to improve prediction accuracy. Figure 5 shows two batches of Ultipa graph-based predictions to significantly improve accuracy (40-50% better) and latency (10-15x) over ML/DL methodologies. The main cause for such improvement is the graph-based feature extraction of the transaction network accurately reflects card holders’ (and merchants’) behavior patterns therefore giving augmented prediction power. In comparison, bigdata/ML predictions are still table centric and lowdimensional which can hardly track the supposedly high-dimensional entity behaviors; besides they are slow and tend to go black-box with sophisticated operations. 2.2
Graph’s natural strength is to be able to analyze data that are connected, and what really sets one graph system apart from the others is its ability to penetrate the data much more deeply within the same time bound and upon the same underpinning hardware.
In Figure 3, we’ve illustrated the necessity for deeptraversal. To implement a graph XAI system with realtime deep traversal capability, there are 3 factors to consider: • • •
Low-latency: data structures that allow for lowest possible access time-complexity, ideally O(1).
Mutability: read-only data structure and access patterns are easier to design but we must cope with read-n-write scenarios where data are mutable and supporting CRUD operations.
Parallelization: serial access per query is easy to do, but the existence of supernodes would require parallel and accelerated access on a single query – because one query can lead to graph-wide traversal.
We could use Map or HashMap in C++ to implement the core meta-data data structure, but both are considered highly redundant in terms of memory consumption. We came up with a novel data-structure design of vector_of_vectors as illustrated in Figure 6, essentially, packing all edges connecting with a vertex in a mutable vector, but aggregating inbound and outbound edges in different sections for easier graph traversal. Such data structure can satisfy the needs for mutability and parallel access, and most importantly O(1) time complexity for per-hop data traversal – the computational complexity to visit all neighbors of a vertex is a constant O(1) which will empower exponentially faster deeper traversals.
There are other forms of novel acceleration
techniques, which we’ll introduce in section 2.4
(Velocity). The effect of deep traversal capability is
shown in Figure 7, where 1-hop traversal with Ultipa is
done in microseconds while the other systems require at
least dozens of milliseconds, and at 6-hop, only 2
systems (Ultipa and Tigergraph) can perform while
Ultipa is nearly 50 times faster than the other. When the
depth reaches 23-hop (this is close to the diameter [
There are many optimizations made in pursuing for real-time recursive deep penetration of dataset. Multilayer storage-n-computing acceleration is one such optimization, which can be reflected in how critical system resources are consumed. Figure 8 shows that Ultipa uses more static memory but less dynamic memory comparing with other systems, meanwhile goes far more parallel in processing graph queries. A salient benefit of lower dynamic memory is equivalent to better system stability while the other systems risk running into OOM with deep-query processing.
G r a p h Q u e r y & R e s o u r c e
C o n s u m p t i o n ( 3 8 4 G B D R A M , 6 4 v C P U )
Graph data is meta-data that are organized and connected in high-dimensional ways, and the finest granularity of meta-data boils down to vertices and edges. By organizing and manipulating these meta-data in different dimensions, insights or certain facets of the graph can be generated on the fly. Explainability often demands for reverse thinking process that is to trace backward from the result to the intermediate process, and eventually to the source data (being analyzed) – if any part of the back-tracing process is hardly explainable, we’ll suspect that the underpinning system has explainability problem even though the system may work well on many aspects.
Visualization is another important aspect that helps with explainability. Figure 3 shows the graphical results of a multiple-hop path finding with different types of entities and relationships clearly annotated for easy digestion. The query language itself is also important, but it would be pointless if we don’t put this in the context of comparing with varied graph query languages. }
Even though readers may be subjective on intuitiveness (explainability) of the above GQL dialects, there are several things that hinder easy comprehension such as non-inline filtering, extensive chaining, or mixing up of SQL and C++ procedural programming styles.
Root of LCR Numerator
Denominator
Explainability can also be reflected in data modeling. Though RDBMS and tables have been main-stream for decades, it can be a disaster to use dozens of tables to serve sophisticated business scenarios like liquidity risk management. Figure 10 shows a novel graph data modeling, where regulated key financial indicator LCR (Liquidity Coverage Ratio) formula is transformed to the graph, intuitively. This enables not only great explainability, but also exponentially accelerated computing of the LCR indicator (due to avoidance of the cartesian-product of joining dozens of tables) with highdensity parallel processing, and attribution analysis which is essentially a back-tracing processing with dynamic filtering on the tree-like graph. 2.4
The velocity aspect of graph XAI is crucial in enabling accelerated deep traversal during big data analytics. There are 3 novel techniques leveraged by Ultipa graph database in accelerating: • • •
High-density graph computing: adapting parallel computing to graph domain with optimizations to penetrate hotspot supernodes in conjunction with native-graph data structures.
Data structure optimization: this has been discussed in the XAI’s depth aspect, essentially vertex-edge adjacency data structures.
Traversal boosting: this encompasses multiple facets, primarily optimized redesign of graph query and algorithm traversal logics, with the help of acceleration data structures.
In Figure 11, a novel graph traversal method is illustrated, instead of traversing from one vertex only, the method allows traversal to be conducted parallelly from both ends. This would exponentially lower the traversal complexity, and the query will return as soon as common neighbors are found in the middle, and the overall theoretical query time-complexity can be exponentially lower. Additionally, in-memory datastructure is used to boost traversal speed (storing temporal neighborhood states). The empirical benchmark data shows that an overall acceleration of 40 times over K-hop neighborhood finding, and 160 time over path-finding are achieved (see Figure 12). As more cores are fired up for denser parallel processing, the acceleration effect is significant (32-core vs. 1-core: 7x for 3-hop, 25x for 6-hop, and 100x for Shortest-path. Note that adding more cores not always yield better performance, in shortest-path finding, 16-core turns out to be slightly faster than 32-core, as communications between more cores tend to add communication costs).
The performance gain of utilizing high-density parallel graph computing and boosted-traversal mechanism can significantly reduce system latency and increase system throughput per QPS and TPS. In realworld applications, Ultipa graph system incorporating these acceleration mechanisms can allow T+1 (1-day) batch processing tasks to be completed in real-time or near-real-time T+0 (same day) fashion, therefore opening up opportunities for broad spectrum business scenario realization (and acceleration).
Ultipa Graph: HDC & Boosting 3. Conclusions and Future Work Big data analysis and AI are undoubtedly integral to our technological landscape. However, several fundamental challenges demand our attention. In this paper, we have demonstrated that explainable AI (XAI) through graph augmentation can provide practical solutions to address key issues (A-D-E-V) in the field.
While significant progress has been made, several avenues remain underexplored. We recommend further investigation in the following areas: • • •
Dynamic Graph Data Modeling: Optimizing on dynamic graph data modeling based on evolving data and temporal user/query requirements.
High-Scalable Graph Computing Architecture: Efficiently handling extra large-scale (zillionscale) graphs.
TP (Transactional Processing) and AP (Analytics Processing) Fusion: Bridging the gap between transactional and analytical workloads.
These research directions hold promise for future breakthroughs, and we anticipate their incorporation into production systems and scholarly publications. Acknowledgements We would like to thank our customers for oefring inspirations and opportunities to realizing their relentless business scenarios with our cutting-edge graph XAI solutions, and big kudos to our teammates Lynsey, Zoey, Bin, Pearl, and countless others for their contribution, inputs, and insightful feedback.