Title:
RANDOM GREEDY ALGORITHM-BASED HORIZONTAL FEDERATED GRADIENT BOOSTED TREE OPTIMIZATION METHOD
Document Type and Number:
WIPO Patent Application WO/2022/151654
Kind Code:
A1
Abstract:
A random greedy algorithm-based horizontal federated gradient boosted tree optimization method, which comprises the following steps: a coordinator configures gradient boosted tree model relevant parameters, which comprise but are not limited to a maximum number of decision trees (T), a maximum tree depth (L), an initial prediction value (base), etc., and sends same to each participant (p_i); and each participant partitions a current node dataset according to a partition feature f and a partition value v, and allocates new partitioned data to a child node. In the present random greedy algorithm-based horizontal federated gradient boosted tree optimization method, horizontal federated learning comprises participants and a coordinator, the participants possess local data, the coordinator does not possess any data, and is a center that aggregates participant information, the participants respectively calculate a histogram and send the histogram to the coordinator, and after collecting all histogram information, the coordinator searches for a maximally optimal partition point according to a greedy algorithm, and then shares same with each participant, and operates in coordination with an internal algorithm.
More Like This:
Inventors:
ZHANG JINYI (CN)
LI ZHENFEI (CN)
LI ZHENFEI (CN)
Application Number:
PCT/CN2021/101319
Publication Date:
July 21, 2022
Filing Date:
June 21, 2021
Export Citation:
Assignee:
ENNEW DIGITAL TECH CO LTD (CN)
International Classes:
G06N3/04; G06K9/62
Foreign References:
| CN109299728A | 2019-02-01 | |||
| CN111985270A | 2020-11-24 | |||
| CN111553483A | 2020-08-18 | |||
| CN111553470A | 2020-08-18 | |||
| US20200200648A1 | 2020-06-25 |
Attorney, Agent or Firm:
BEIJING JIAKE INTELLECTUAL PROPERTY LAW FIRM (CN)
Download PDF: