cleora
Cleora AI is a general-purpose open-source model for efficient, scalable learning of stable and inductive entity embeddings for heterogeneous relational data. C
What is cleora?
Cleora AI is a general-purpose open-source model for efficient, scalable learning of stable and inductive entity embeddings for heterogeneous relational data. Created by Synerise.com team.
Overview of cleora
What it is
Cleora is an open-source model for learning entity embeddings from heterogeneous relational data. It is developed by the Synerise.com team. The software turns a graph or hypergraph into one vector per node by repeatedly normalizing the average of each node's neighbors' vectors. It runs on a CPU without sampling, training, or a GPU.
Key features
- Fast: embeds ogbn-arxiv (169,343 nodes, 1.2 million edges) at 1024 dimensions in 1.4 seconds on a laptop CPU.
- Accurate: ranks first among ten methods by mean rank in similarity search, node classification, and link ranking on most graphs tested.
- Lean: uses only two embedding matrices, requiring a fifth to a tenth of the memory of other fast methods.
- Scales: graphs with tens of millions of edges are processed in minutes.
- Deterministic: same input yields identical embeddings regardless of thread count.
- Simple: no negative sampling, learning rate, epochs, or GPU required; the main parameter is the number of iterations.
- Practical: supports hypergraphs, weights, bipartite data, new nodes without retraining, and user-provided feature vectors.
Use cases
Cleora is used for similarity search, node classification, and link ranking on relational data. It can embed graphs and hypergraphs from domains such as e-commerce (product baskets), social networks, and recommendation systems. The software is also applicable to bipartite relations (e.g., user–product) and can incorporate popularity priors for link prediction.
Getting started
bash pip install pycleora Python 3.10 to 3.15 on Linux (x86-64, arm64), macOS (Apple silicon, Intel), and Windows. NumPy is the only required dependency. python from pycleora import SparseMatrix, embed, find_most_similar baskets = [ 'laptop mouse keyboard', 'laptop mouse monitor', 'keyboard mouse', 'espresso_machine coffee_beans', 'coffee_beans mug', 'espresso_machine coffee_beans mug', ] graph = SparseMatrix.from_iterator(iter(baskets), 'complex::reflexive::product') embeddings = embed(graph, feature_dim=64, num_iterations=2) for hit in find_most_similar(graph, embeddings, 'laptop', top_k=3): print(f'{hit['entity_id']}: {hit['similarity']:.3f}') For a plain graph, pass edges as arrays: python import numpy as np source = np.array([0, 0, 1, 2, 3]) target = np.array([1, 2, 2, 3, 4]) network = SparseMatrix.from_edges(source, target, columns='complex::reflexive::node') vectors = embed(network, feature_dim=1024)
License and pricing
The license is not stated in the provided metadata or README. The software is categorized as open source.
這是什麼
Cleora 是一個用於從異質關聯資料學習實體嵌入的開源模型,由 Synerise.com 團隊開發。該軟體透過重複對每個節點鄰居的向量進行平均並正規化,將圖形或超圖轉換為每個節點一個向量。它在 CPU 上運行,無需取樣、訓練或 GPU。
主要功能
- 快速:在筆記型電腦 CPU 上,以 1.4 秒將 ogbn-arxiv(169,343 個節點,120 萬條邊)嵌入至 1024 維度。
- 準確:在相似度搜尋、節點分類和連結排序中,於大多數測試圖形上,在十種方法中平均排名第一。
- 精簡:僅使用兩個嵌入矩陣,記憶體需求僅為其他快速方法的五分之一到十分之一。
- 可擴展:數千萬條邊的圖形可在數分鐘內處理完成。
- 確定性:無論執行緒數量為何,相同的輸入都會產生完全相同的嵌入。
- 簡單:無需負取樣、學習率、訓練週期或 GPU;主要參數為迭代次數。
- 實用:支援超圖、權重、二分資料、無需重新訓練的新節點,以及使用者提供的特徵向量。
使用情境
Cleora 用於關聯資料的相似度搜尋、節點分類和連結排序。它可以嵌入來自電子商務(產品購物籃)、社交網路和推薦系統等領域的圖形和超圖。該軟體也適用於二分關係(例如使用者–產品),並可結合流行度先驗進行連結預測。
快速開始
bash pip install pycleora Python 3.10 至 3.15,支援 Linux(x86-64, arm64)、macOS(Apple silicon, Intel)和 Windows。NumPy 是唯一必需的依賴項。 python from pycleora import SparseMatrix, embed, find_most_similar baskets = [ 'laptop mouse keyboard', 'laptop mouse monitor', 'keyboard mouse', 'espresso_machine coffee_beans', 'coffee_beans mug', 'espresso_machine coffee_beans mug', ] graph = SparseMatrix.from_iterator(iter(baskets), 'complex::reflexive::product') embeddings = embed(graph, feature_dim=64, num_iterations=2) for hit in find_most_similar(graph, embeddings, 'laptop', top_k=3): print(f'{hit['entity_id']}: {hit['similarity']:.3f}') 對於一般圖形,以陣列形式傳遞邊: python import numpy as np source = np.array([0, 0, 1, 2, 3]) target = np.array([1, 2, 2, 3, 4]) network = SparseMatrix.from_edges(source, target, columns='complex::reflexive::node') vectors = embed(network, feature_dim=1024)
授權與價格
提供的元資料或 README 中未說明授權條款。該軟體被歸類為開源。
Based on the project's public README and repository metadata.