About Expander working principle partAn excerpt from Google Research, the author is Ravi Sujith.
Recently, the Google Research Expander team released a large, graph based machine learning platform, and this technique is used in our daily life, to remind the inbox Allo intelligent information reply, Google Photos image recognition is one of the strong support behind the function.
So what exactly is the semi supervised algorithm based on graph? How is it working? And with our financial anti fraud identification of what is the relationship and role? The following small for everyone to look at the science of krypton:
What is Learning Graph?
How to construct a graph in Expander?
Graph based semi supervised algorithm is the core of the construction of the graph itself. So how to build a picture? Need to define the node, edge, as well as the edge of the weight (edge of the weight that is the similarity of the node). In Natural Language Processing, for example, in the case of emotional analysis, each node represents a section of text information, and those edges is the similarity of the text emotion.
How is it working?
At its core, Expander platform, combined with semi supervised learning and large-scale graph based learning to construct a multi data representation. For example, as shown in the figure, there are two types of nodes: gray represents no label data, color on behalf of the label data. The relationship between the nodes is expressed by the edge, and the strength of the connection is expressed by the edge of the edge. We now define the goal of semi supervised learning: the color of each node in the forecast map.
Secondly, Expander graph learning framework is used to solve this task as an optimization problem. At the simplest level, it first studies the color tag of each node in the graph, and then assigns the color to the adjacent nodes according to the strength of the connection. A naive way is the color of the label distribution try to finish all disposable nodes, but it cannot be extended to large map. Therefore, the better solution is: the color of the node to the adjacent nodes, and then repeat the process. As shown in the figure, in each step, by observing the color distribution of neighboring nodes, a label free node can be assigned a label. In this way we can update the label of each node, repeat the operation until the entire graph is colored, and this process is proved to be very effective in the optimization of similar problems.
How to apply the graph based semi supervised learning in the financial scene?
Krypton letter network anti fraud services interface example
The analogy above Graph Learning, in the financial scene, in fact, every applicant, mobile phone number, equipment, IP addresses are nodes in the graph, such as the applicant owned equipment, mobile phone number mobile phone number to call contact is the side of the figure, the edge weight is closely related. In building a map, the applicant default or not mark is the original seed node, by using a semi supervised algorithm based on the graph, the label propagation will default to the applicant without a label, so that we can in a small amount of labeled samples is constructed on the huge risk of network, and ultimately create a form our reliable default prediction model.
It has a number of cooperation letter of large financial institutions to explore mature application of semi supervised algorithm based on graph control in the field of financial risk, as shown above, the first is related to the risk of simplified network, light green, yellow and red nodes represent 1 (low risk), 2 (stroke, 3 (insurance) high risk; second) for global risk network, on behalf of the grey node expansion 0 (no risk). For each applicant, cell phone number, etc., can be based on the map database network anti fraud services, given the appropriate level of group risk of fraud and detailed risk description.
The letter is krypton through advanced ID Mapping and fuzzy matching technology to help customers build financial, personal or business complex relationship network, XCloud also launched a cloud risk analysis of SNA social network fraud and related risk groups based on strong screening service, for the customer to complete the upgrade from the individual risk to global network risk recognition ability.