In [1]:
# code for loading the format for the notebook
import os
# path : store the current path to convert back to it later
path = os.getcwd()
os.chdir(os.path.join('..', '..', 'notebook_format'))
from formats import load_style
load_style(css_style='custom2.css', plot_style=False)
Out[1]:
In [2]:
os.chdir(path)
# 1. magic for inline plot
# 2. magic to print version
# 3. magic so that the notebook will reload external python modules
# 4. magic to enable retina (high resolution) plots
# https://gist.github.com/minrk/3301035
%matplotlib inline
%load_ext watermark
%load_ext autoreload
%autoreload 2
%config InlineBackend.figure_format='retina'
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
%watermark -a 'Ethen' -d -t -v -p numpy,pandas,sklearn,matplotlib
Similarity search is a widely used and important method in many applications. One example is Shazam, the app that let's us identify can song within seconds is leveraging audio fingerprinting and most likely a fast and scalable similarity search method to retrieve the relevant song from a massive database of songs. In this documentation, we'll be introducing Locality Sensitive Hashing (LSH), an approximate nearest neighborhood search technique in the context of recommendation system. Note that, Locality Sensitive Hashing (LSH) is actually a family of algorithm, different distance metric will correspond to a different method. Here we will be focusing on cosine distance.
In the realm of recommender system, collaborative filtering methods such as matrix factorization are often times workhorse methods of choice. The main idea behind collaborative filtering is that the product you are most likely to buy, is the product that a bunch of people like you also bought. The input to the algorithm is basically a matrix of user IDs and product IDs with values representing who bought what product. Albeit an extremely powerful method, it does suffer from what is known as the cold-start problem.
Let's say we want to make recommendations to a customer viewing a product details page, who just got there from a Google SERP link. We have no historical record about this customer, so we can't build a matrix of purchases. This is where a different approach, so called content-based filtering/recommendation can be helpful.
When we say content-based filtering, we are referring to the fact that we're using items or users' meta-data to generate the recommendation. For example, say the product page that this user landed on is "Nike Pro Hypercool Fitted Men's Compression Shirt", then we can leverage information such as the item's description, price or other attribute to generate recommendation saying you might also love the item "Nike Pro Hypercool Printed Men's Tights".
The dataset that we'll be working with here is a sample data consisting of outdoor clothing and products from Patagonia. It only consists of two columns, the IDs and the product's text description. To leverage this information, we will transform the raw text into a tfidf-format, and use the tf-idf representation to find similar items for a given item (the most similar items to the item we're interested in is essentially our recommended items).
In [3]:
input_local_path = os.path.join('data', 'sample-data.csv')
df = pd.read_csv(input_local_path)
print('dimension: ', df.shape)
df.head()
Out[3]:
In [4]:
from sklearn.feature_extraction.text import TfidfVectorizer
tfidf = TfidfVectorizer(
analyzer='word',
ngram_range=(1, 3),
min_df=0,
stop_words='english')
X_tfidf = tfidf.fit_transform(df['description'])
X_tfidf
Out[4]:
In [5]:
def get_similarity_items(X_tfidf, item_id, topn=5):
"""
Get the top similar items for a given item id.
The similarity measure here is based on cosine distance.
"""
query = X_tfidf[item_id]
scores = X_tfidf.dot(query.T).toarray().ravel()
best = np.argpartition(scores, -topn)[-topn:]
return sorted(zip(best, scores[best]), key=lambda x: -x[1])
similar_items = get_similarity_items(X_tfidf, item_id=1)
# an item is always most similar to itself, in real-world
# scenario we might want to filter itself out from the output
for similar_item, similarity in similar_items:
item_description = df.loc[similar_item, 'description']
print('similar item id: ', similar_item)
print('cosine similarity: ', similarity)
print('item description: ', item_description)
print()
Hopefully, we can sort of agree that the result reflects our intuition. Given a item, Active sport boxer briefs, the content-based filtering method based on the item's description in tfidf-representation recommends items such as Active sport briefs, Cap 1 boxer briefs which are all underwear-related products.
To generate the similar items, the approach we're using here is calculating the cosine distance between our query item and all the other items in our data, then sorting the distance to find the most similar items. This approach might work well for a small dataset, however, as we can imagine this procedure might become a bottleneck to our system as the data points starts to increase, and in near-real time systems where there's a strict latency requirements, this method is most likely not going to cut it.
The idea behind LSH is to throw down random lines/vectors. Yes, you are reading it correctly. Then everything that falls below this line has a negative score and will fall into what we'll be referring to as bin 0. On the other hand, everything that's above this line has a positive score and all of those points will be assigned to bin 1. So we're essentially translating the sign of the scores into a binary index.
For example, in the picture above, the orange data point falls above the random vector and we'll label it as a white bin. Then we can use this for nearest neighbor search in the following way: Given a query data point, say it also falls above the random vector, then we will only search for the nearest neighbor amongst the data points that also fall above the random vector, or another way to put it, we will only search amongst the data points that also fall under the same bin.
If we think about this carefully, the main rationale behind locality sensitive hashing is data points that are located close to each other should be mapped to similar hashes (in same bin/bucket with high probability). i.e. in our made-up example, data points that are close to the orange data point should fall in the same bin, whereas data points that are far away should fall in different bin. However, because of the randomness, it is not likely that all similar items are grouped correctly. Hence, to overcome this limitation, a common practice is to create multiple random vectors.
e.g. Say we throw down another random vector, this time the orange data point falls under this random vector, thus it falls under a different bin (black bin instead of white).
Since, each bin is represented by a bitwise hash value, which is a number made up of a sequence of 1s and 0s. e.g. for this orange data point, if we consider the white bin as 0 and the black bin as 1, then this data point's bin will be represented by [0, 1].
Let's use some code to illustrate this process. Our first step is to generate a collection of random vectors from the standard Gaussian distribution.
In [6]:
def generate_random_vectors(dim, n_vectors):
"""
generate random projection vectors
the dims comes first in the matrix's shape,
so we can use it for matrix multiplication.
"""
return np.random.randn(dim, n_vectors)
We now generate random vectors of the same dimensionality as our vocabulary size. Each vector can be used to compute one bit in the bin encoding. We generate 16 vectors, leading to a 16-bit encoding of the bin index for each document. Note that 16 is a hyperparamter, we will look into this adjustable parameter in later section.
In [7]:
vocab_size = len(tfidf.get_feature_names())
print('vocabulary size: ', vocab_size)
np.random.seed(0)
n_vectors = 16
random_vectors = generate_random_vectors(vocab_size, n_vectors)
print('dimension: ', random_vectors.shape)
random_vectors
Out[7]:
Next, we'd like to decide which bin each documents should go. Since 16 random vectors were generated in the previous cell, we have 16 bits to represent the bin index. The first bit is given by the sign of the dot product between the first random vector and the document's TF-IDF vector, and so on.
In [8]:
# use one data point's tfidf representation as an example
data_point = X_tfidf[0]
# True if positive sign; False if negative sign
bin_indices_bits = data_point.dot(random_vectors) >= 0
print('dimension: ', bin_indices_bits.shape)
bin_indices_bits
Out[8]:
All documents that obtain exactly this vector will be assigned to the same bin. One further preprocessing step we'll perform is to convert this resulting bin into integer representation. This makes it convenient for us to refer to individual bins.
Bin index integer
[0,0,0,0,0,0,0,0,0,0,0,0] => 0
[0,0,0,0,0,0,0,0,0,0,0,1] => 1
[0,0,0,0,0,0,0,0,0,0,1,0] => 2
[0,0,0,0,0,0,0,0,0,0,1,1] => 3
...
[1,1,1,1,1,1,1,1,1,1,0,0] => 65532
[1,1,1,1,1,1,1,1,1,1,0,1] => 65533
[1,1,1,1,1,1,1,1,1,1,1,0] => 65534
[1,1,1,1,1,1,1,1,1,1,1,1] => 65535 (= 2^16-1)By the rules of binary number representation, to perform the conversion, we can compute the dot product between the document vector and the vector consisting of powers of 2:
In [9]:
bin_indices_bits = data_point.dot(random_vectors) >= 0
# https://wiki.python.org/moin/BitwiseOperators
# x << y is the same as multiplying x by 2 ** y
powers_of_two = 1 << np.arange(n_vectors - 1, -1, step=-1)
print(powers_of_two)
# final integer representation of individual bins
bin_indices = bin_indices_bits.dot(powers_of_two)
print(bin_indices)
We can repeat the identical operation on all documents in our dataset and compute the corresponding bin. We'll again leverage matrix operations so that no explicit loop is needed.
In [10]:
# we can do it for the entire corpus
bin_indices_bits = X_tfidf.dot(random_vectors) >= 0
print(bin_indices_bits.shape)
bin_indices = bin_indices_bits.dot(powers_of_two)
bin_indices.shape
Out[10]:
The resulting array gives us the integer index of the bins for all documents.
Now we are ready to complete the following function. Given the integer bin indices for the documents, we would curate the list of document IDs that belong to each bin. Since a list is to be maintained for each unique bin index, a dictionary of lists is used.
In [11]:
from collections import defaultdict
def train_lsh(X_tfidf, n_vectors, seed=None):
if seed is not None:
np.random.seed(seed)
dim = X_tfidf.shape[1]
random_vectors = generate_random_vectors(dim, n_vectors)
# partition data points into bins,
# and encode bin index bits into integers
bin_indices_bits = X_tfidf.dot(random_vectors) >= 0
powers_of_two = 1 << np.arange(n_vectors - 1, -1, step=-1)
bin_indices = bin_indices_bits.dot(powers_of_two)
# update `table` so that `table[i]` is the list of document ids with bin index equal to i
table = defaultdict(list)
for idx, bin_index in enumerate(bin_indices):
table[bin_index].append(idx)
# note that we're storing the bin_indices here
# so we can do some ad-hoc checking with it,
# this isn't actually required
model = {'table': table,
'random_vectors': random_vectors,
'bin_indices': bin_indices,
'bin_indices_bits': bin_indices_bits}
return model
# train the model
n_vectors = 16
model = train_lsh(X_tfidf, n_vectors, seed=143)
After generating our LSH model, let's examine the generated bins to get a deeper understanding of them. Recall that during the background section, given a product's tfidf vector representation, we were able to find its similar product using standard cosine similarity. Here, we will look at these similar products' bins to see if the result matches intuition. Remember the idea behind LSH is that similar data points will tend to fall into nearby bins.
In [12]:
# comparison
similar_item_ids = [similar_item for similar_item, _ in similar_items]
bits1 = model['bin_indices_bits'][similar_item_ids[0]]
bits2 = model['bin_indices_bits'][similar_item_ids[1]]
print('bits 1: ', bits1)
print('bits 2: ', bits2)
print('Number of agreed bins: ', np.sum(bits1 == bits2))
In [13]:
# comparison
bits1 = model['bin_indices_bits'][similar_item_ids[0]]
bits2 = model['bin_indices_bits'][similar_item_ids[4]]
print('bits 1: ', bits1)
print('bits 2: ', bits2)
print('Number of agreed bins: ', np.sum(bits1 == bits2))
Looking at the result above, it does seem like LSH is doing what its intended to do: i.e., similar data points will agree upon more bin indices than dissimilar data points, however, in a high-dimensional space such as text features, we can get unlucky with our selection of only a few random vectors such that dissimilar data points go into the same bin while similar data points fall into different bins. Hence, given a query document, we should consider all documents in the nearby bins and sort them according to their actual distances from the query.
Let's first implement the logic for searching nearby neighbors, which goes like this:
To obtain candidate bins that differ from the query bin by some number of bits, we use itertools.combinations, which produces all possible subsets of a given list. See this documentation for details.
In [14]:
from itertools import combinations
def search_nearby_bins(query_bin_bits, table, search_radius=3, candidate_set=None):
"""
For a given query vector and trained LSH model's table
return all candidate neighbors with the specified search radius.
Example
-------
model = train_lsh(X_tfidf, n_vectors=16, seed=143)
query = model['bin_index_bits'][0] # vector for the first document
candidates = search_nearby_bins(query, model['table'])
"""
if candidate_set is None:
candidate_set = set()
n_vectors = query_bin_bits.shape[0]
powers_of_two = 1 << np.arange(n_vectors - 1, -1, step=-1)
for different_bits in combinations(range(n_vectors), search_radius):
# flip the bits (n_1, n_2, ..., n_r) of the query bin to produce a new bit vector
index = list(different_bits)
alternate_bits = query_bin_bits.copy()
alternate_bits[index] = np.logical_not(alternate_bits[index])
# convert the new bit vector to an integer index
nearby_bin = alternate_bits.dot(powers_of_two)
# fetch the list of documents belonging to
# the bin indexed by the new bit vector,
# then add those documents to candidate_set;
# make sure that the bin exists in the table
if nearby_bin in table:
candidate_set.update(table[nearby_bin])
return candidate_set
The next code chunk uses this searching for nearby bins logic to search for similar documents and return a dataframe that contains the most similar data points according to LSH and their corresponding distances.
In [15]:
from sklearn.metrics.pairwise import pairwise_distances
def get_nearest_neighbors(X_tfidf, query_vector, model, max_search_radius=3):
table = model['table']
random_vectors = model['random_vectors']
# compute bin index for the query vector, in bit representation.
bin_index_bits = np.ravel(query_vector.dot(random_vectors) >= 0)
# search nearby bins and collect candidates
candidate_set = set()
for search_radius in range(max_search_radius + 1):
candidate_set = search_nearby_bins(bin_index_bits, table, search_radius, candidate_set)
# sort candidates by their true distances from the query
candidate_list = list(candidate_set)
candidates = X_tfidf[candidate_list]
distance = pairwise_distances(candidates, query_vector, metric='cosine').flatten()
distance_col = 'distance'
nearest_neighbors = pd.DataFrame({
'id': candidate_list, distance_col: distance
}).sort_values(distance_col).reset_index(drop=True)
return nearest_neighbors
In [16]:
print('original similar items:\n' + str(similar_items))
item_id = 1
query_vector = X_tfidf[item_id]
nearest_neighbors = get_nearest_neighbors(X_tfidf, query_vector, model, max_search_radius=5)
print('dimension: ', nearest_neighbors.shape)
nearest_neighbors.head()
Out[16]:
We can observe from the result above that when we use a max_search_radius of 5, our LSH-based similarity search wasn't capable of retrieving the actual most similar items to our target data point, this is expected as we have mentioned LSH is an approximate nearest neighborhood search method. However, if we were to increase the radius parameter to 10, we can in fact retrieve all the actual most similar items.
In [17]:
nearest_neighbors = get_nearest_neighbors(X_tfidf, query_vector, model, max_search_radius=10)
print('dimension: ', nearest_neighbors.shape)
nearest_neighbors.head()
Out[17]:
In [18]:
# we can perform a join with the original table to get the description
# for sanity checking purpose
nearest_neighbors.head().merge(df, on='id', how='inner')
Out[18]:
In the following sections we have implemented a few experiments so we can gain some intuition as to how LSH behaves in different situations. This will help us understand the effect of searching nearby bins and the performance of LSH versus computing nearest neighbors using a brute force search.
The first experiment that we'll be looking at is the effect of nearby bin search.
How does nearby bin search affect the three variables that are tied to the search radius:
Let's run LSH multiple times, each with different radius for nearby bin search. We will keep track of the three variables that were discussed above.
In [19]:
import time
query_time_history = []
num_candidates_history = []
max_distance_history = []
min_distance_history = []
average_distance_history = []
topn = 5
query_vector = X_tfidf[1]
for max_search_radius in range(12):
start = time.time()
nearest_neighbors = get_nearest_neighbors(X_tfidf, query_vector, model, max_search_radius)
end = time.time()
query_time = end - start
# the closest neighbor is the query point itself, thus
# we'll exclude it from the calculation
distances = nearest_neighbors['distance'].iloc[1:].head(topn)
max_distance = distances.max()
min_distance = distances.min()
average_distance = distances.mean()
num_candidates = nearest_neighbors.shape[0]
query_time_history.append(query_time)
num_candidates_history.append(num_candidates)
max_distance_history.append(max_distance)
min_distance_history.append(min_distance)
average_distance_history.append(average_distance)
In [20]:
# change default style figure and font size
plt.rcParams['figure.figsize'] = 8, 6
plt.rcParams['font.size'] = 12
linewidth = 4
plt.figure()
plt.plot(num_candidates_history, linewidth=4)
plt.xlabel('Search radius')
plt.ylabel('# of documents searched')
plt.tight_layout()
plt.figure()
plt.plot(query_time_history, linewidth=linewidth)
plt.xlabel('Search radius')
plt.ylabel('Query time (seconds)')
plt.tight_layout()
plt.figure()
plt.plot(average_distance_history, linewidth=linewidth, label='Average of 10 neighbors')
plt.plot(max_distance_history, linewidth=linewidth, label='Farthest of 10 neighbors')
plt.plot(min_distance_history, linewidth=linewidth, label='Closest of 10 neighbors')
plt.xlabel('Search radius')
plt.ylabel('Cosine distance of neighbors')
plt.legend(loc='best')
plt.tight_layout()
Some observations:
The next experiment that we'll be conducting is one that we are probably most interested in: Quality metrics for neighbors. Since LSH is an approximate nearest neighborhood method, often times, we would like to have a metric such as precision or recall to measure the quality of the method. For example, with precision, we would be able to answer questions such as: How many of the 10 neighbors given by LSH are among the true 10 nearest neighbors?
Our previous experiment is limited by the fact that it was run with a single query (data point). We should repeat the analysis for the entire data to make sure it generalizes. As Iterating over all data points would take a long time, we will cheat here and randomly choose a small subset.
In [21]:
topn = 10
max_radius = 11
num_queries = 10
np.random.seed(0)
# key = radius, value = list of precision for different queries
precision_history = {i: [] for i in range(max_radius)}
random_query_ids = np.random.choice(X_tfidf.shape[0], num_queries, replace=False)
for i, query_id in enumerate(random_query_ids):
print('%s / %s' % (i + 1, num_queries))
# get the set of true nearest neighbors
similar_items = get_similarity_items(X_tfidf, item_id=query_id, topn=topn)
ground_truth = set(similar_item for similar_item, _ in similar_items)
# compute precision metric for each radius
query_vector = X_tfidf[query_id]
for radius in range(max_radius):
nearest_neighbors = get_nearest_neighbors(X_tfidf, query_vector, model, max_search_radius=radius)
candidate_set = set(nearest_neighbors['id'].head(topn))
# precision = (# of neighbors both in result and ground_truth)
precision = len(candidate_set & ground_truth) / topn
precision_history[radius].append(precision)
precision_history
Out[21]:
In [22]:
mean_precision = [np.mean(precision_history[i]) for i in range(len(precision_history))]
plt.figure()
plt.plot(mean_precision, linewidth=4, label='Precison@' + str(topn))
plt.xlabel('Search radius')
plt.ylabel('Precision')
plt.legend()
plt.tight_layout()
In this article, we saw that LSH performs an efficient neighbor search by randomly partitioning all reference data points into different bins, when it comes to the similarity search stage, it will only consider searching within data points that fall within the same bin. Then a radius parameter gives the end-user full control over the trade-off between the speed of the search versus the quality of the nearest neighbors.
There are many other applications of LSH, here we are using it with text-based features, the same idea can also be applied to images as well. And for those interested, the following link contains a blog post of its use-case at Uber, where they are using the technique to find similar trips based on their spatial properties, a method for identifying fraudulent drivers. Blog: Detecting Abuse at Scale: Locality Sensitive Hashing at Uber Engineering