Clustering is an unsupervised learning technique that groups items by similarity. In search, it turns embedding vectors for products or documents into data-driven categories that can guide personalization and relevance decisions. section 9.4.2 section 1.1
A practical pipeline embeds items, fits a clustering algorithm, assigns each item a label, and inspects representative terms to check whether the groups form coherent categories. PCA can reduce cluster-center embeddings for a two-dimensional diagnostic plot, but the visualization is for inspection rather than clustering itself. section 9.4.2
Dynamic product categories
A catalog of nearly 50,000 products is grouped into 100 KMeans clusters. Inspecting top words reveals coherent areas such as kitchen appliances, music, video, and data storage. A query such as “microwave” maps to the microwave cluster; comparing nearest cluster centers can also return related kitchen-product clusters. section 9.4.2
Listing 9.13 Generating dynamic categories from clustered products
def get_clusters(data, algorithm, args):
return algorithm(**args).fit(data)
def assign_clusters(labels, product_names):
clusters = defaultdict(lambda:[], {})
for i in range(len(labels)):
clusters[labels[i]].append(product_names[i])
return clusters
args = {"n_clusters": 100, "n_init": 10, "random_state": 0} #1
algo = get_clusters(product_embeddings, cluster.KMeans, args) #1
labels = algo.predict(product_embeddings)
clusters = assign_clusters(labels, product_names) #2
Shows the compact Python pattern for fitting KMeans, predicting labels, and collecting product names into dynamic categories. section 9.4.2
Listing 9.15 Comparing techniques for mapping queries to clusters
import sentence_transformers, heapq
def get_top_labels_centers(query, centers, n=2): #1
query_embedding = transformer.encode([query], convert_to_tensor=False)
similarities = sentence_transformers.util.cos_sim(
query_embedding, centers)
sim = similarities.tolist()[0]
return [sim.index(i) for i in heapq.nlargest(n, sim)]
def get_query_cluster(query): #2
query_embedding = transformer.encode([query], convert_to_tensor=False)
return algo.predict(query_embedding)
def get_cluster_description(cluser_num):
return "_".join(top_words(clusters[cluser_num], 5))
query = "microwave"
kmeans_predict = get_query_cluster(query)[0] #3
print("K-means Predicted Cluster:")
print(f" {kmeans_predict} ({get_cluster_description(kmeans_predict)})")
closest_sim = get_top_labels_centers(query, centers, 1)[0] #4
print(f"\nCosine Predicted Cluster:")
print(f" {closest_sim} ({get_cluster_description(closest_sim)})")
knn_cosine_similarity = get_top_labels_centers(query, centers, 5) #5
print(f"\nKNN Cosine Predicted Clusters: {knn_cosine_similarity}")
for n in knn_cosine_similarity:
print(f" {n} ({get_cluster_description(n)})")
Compares assigning a query through KMeans prediction, similarity to cluster centers, and nearest-neighbor cosine similarity. section 9.4.2
