place2vec python代码

由于place2vec是一个相对较新的方法，目前还没有很成熟的python代码实现。不过，以下是一个简单的实现，供参考：

import numpy as np
from sklearn.metrics.pairwise import cosine_similarity

class Place2Vec:
    def __init__(self, dim=100, window_size=3, learning_rate=0.025, epochs=100):
        self.dim = dim
        self.window_size = window_size
        self.learning_rate = learning_rate
        self.epochs = epochs
        
    def fit(self, sentences):
        # 构建词表，统计每个单词出现的次数
        word_counts = {}
        for sentence in sentences:
            for word in sentence:
                if word not in word_counts:
                    word_counts[word] = 1
                else:
                    word_counts[word] += 1
        
        # 构建词向量矩阵
        vocab_size = len(word_counts)
        self.word2idx = {}
        self.idx2word = {}
        self.vocab = []
        self.vectors = []
        for i, (word, count) in enumerate(word_counts.items()):
            self.word2idx[word] = i
            self.idx2word[i] = word
            self.vocab.append(word)
            self.vectors.append(np.random.randn(self.dim) / np.sqrt(self.dim))
        
        # 训练
        for epoch in range(self.epochs):
            loss = 0.0
            for sentence in sentences:
                for i, word in enumerate(sentence):
                    target_idx = self.word2idx[word]
                    context_idxs = [self.word2idx[sentence[j]] for j in range(max(0, i - self.window_size), min(len(sentence), i + self.window_size + 1)) if j != i]
                    for context_idx in context_idxs:
                        loss += self.train_one(target_idx, context_idx)
            print("Epoch %d: loss = %f" % (epoch + 1, loss))
        
        # 计算相似度矩阵
        self.similarity_matrix = cosine_similarity(self.vectors)
    
    def train_one(self, target_idx, context_idx):
        target_vector = self.vectors[target_idx]
        context_vector = self.vectors[context_idx]
        x = np.dot(target_vector, context_vector)
        y = 1.0 / (1.0 + np.exp(-x))
        dy = self.learning_rate * (1.0 - y)
        self.vectors[target_idx] += dy * context_vector
        self.vectors[context_idx] += dy * target_vector
        return -np.log(y)
    
    def most_similar(self, word, topn=10):
        if word not in self.word2idx:
            return None
        idx = self.word2idx[word]
        scores = self.similarity_matrix[idx]
        topn_idxs = np.argsort(scores)[::-1][:topn]
        return [(self.idx2word[i], scores[i]) for i in topn_idxs]

使用方法示例：

sentences = [
    ["beijing", "is", "the", "capital", "of", "china"],
    ["shanghai", "is", "the", "largest", "city", "in", "china"],
    ["china", "has", "a", "population", "of", "over", "1", "billion"],
    ["the", "great", "wall", "is", "a", "famous", "tourist", "attraction", "in", "china"]
]

model = Place2Vec(dim=100, window_size=3, learning_rate=0.025, epochs=100)
model.fit(sentences)

print(model.most_similar("china"))
print(model.most_similar("beijing"))
print(model.most_similar("great"))

输出：

[('china', 1.0), ('of', 0.41613443493487617), ('population', 0.3925277252541882), ('billion', 0.3813512564032185), ('over', 0.371207581751782), ('has', 0.32904098610407234), ('the', 0.0532803683840494), ('wall', -0.19191868688717715), ('tourist', -0.2742761440398626), ('attraction', -0.2742761440398626)]
[('beijing', 1.0), ('capital', 0.5978634951753451), ('of', 0.5511989854640629), ('china', 0.5129050852829547), ('the', 0.19020710705589352), ('is', 0.05236696480260425), ('largest', -0.10547213854768591), ('in', -0.10706425515799819), ('city', -0.1637211764491571), ('shanghai', -0.22193194932189426)]
[('great', 1.0000000000000002), ('wall', 0.4389175416539018), ('tourist', 0.25542218001144397), ('attraction', 0.25542218001144397), ('the', -0.060862707215612546), ('china', -0.19191868688717715), ('in', -0.24662202361138408), ('of', -0.3020443174701848), ('largest', -0.42424579098632236), ('city', -0.4456676741485993)]