Made by Xinyu Chen โข ๐ https://xinychen.github.io
Uber movement project provides data and tools for cities to more deeply understand and address urban transportation problems and challenges. Uber movement speed data measure hourly street speeds across a city (e.g., New York City, Seattle, and London) to enable data-driven city planning and decision making. These data are indeed multivariate time series with N road segments and T time steps (hours), and are featured as high-dimensional, sparse, and nonstationary. To overcome the challenge created by these complicated data behaviors, we propose a temporal matrix factorization framework for multivariate time series forecasting on high-dimensional and sparse Uber movement speed data.
A detailed introduction to the analysis of missing data problem in Uber movement speed data is available on Medium.
- Open the download page of Uber movement project. Take an example of New York City, please try to open NYC Uber movement speed data.
- Set the product as
speeds
and one specific time period. - Download data and save it on your computer.
Figure 1: Uber movement speed heatmaps of New York City (left panel) and Seattle (right panel), USA.
Please download movement-speeds-hourly-new-york-2019-1.csv
(movement speed data file of New York City in January 2019).
import pandas as pd
import numpy as np
data = pd.read_csv('movement-speeds-hourly-new-york-2019-1.csv')
road = data.drop_duplicates(['osm_way_id', 'osm_start_node_id', 'osm_end_node_id'])
road = road.drop(['year', 'month', 'day', 'hour', 'utc_timestamp', 'segment_id', 'start_junction_id',
'end_junction_id', 'speed_mph_mean', 'speed_mph_stddev'], axis = 1)
road.to_csv('road.csv')
In New York City, Uber movement project covers 98,210 road segments.
This process is time-consuming.
import numpy as np
import pandas as pd
month = 1
data = pd.read_csv('movement-speeds-hourly-new-york-2019-{}.csv'.format(month))
road = pd.read_csv('road.csv')
tensor = np.zeros((road.shape[0], max(data.day.values), 24))
k = 0
for i in range(road.shape[0]):
temp = data[(data['osm_way_id'] == road.osm_way_id.iloc[i])
& (data['osm_start_node_id'] == road.osm_start_node_id.iloc[i])
& (data['osm_end_node_id'] == road.osm_end_node_id.iloc[i])]
for j in range(temp.shape[0]):
tensor[k, temp.day.iloc[j] - 1, temp.hour.iloc[j]] = temp.speed_mph_mean.iloc[j]
k += 1
if (k % 1000) == 0:
print(k)
mat = tensor.reshape([road.shape[0], max(data.day.values) * 24])
np.savez_compressed('hourly_speed_mat_2019_{}.npz'.format(month), mat)
del data, tensor
The matrix's row corresponds to one specific road segment, while the column corresponds to one specific hour.
In this repository, we prepare the dataset and place it at the folder datasets/NYC-movement-data-set
:
hourly_speed_mat_2019_1.npz
(91 MB): data is of size 98,210 x 744 with 23,228,581 positive speed observations.hourly_speed_mat_2019_2.npz
(85.2 MB): data is of size 98,210 x 672 with 21,912,460 positive speed observations.hourly_speed_mat_2019_3.npz
(38.1 MB): data is of size 98,210 x 264 with 10,026,045 positive speed observations.
Note that to make the data as small as possible, we only maintain the data during the first 11 days of March 2019, and save it as hourly_speed_mat_2019_3.npz
. You can use the following code to drop the unnecessary portion when preprocessing the raw data.
month = 3
data = pd.read_csv('movement-speeds-hourly-new-york-2019-{}.csv'.format(month))
road = pd.read_csv('road.csv')
i = data[(data.day > 11)].index
data = data.drop(i)
You can also consider to use the prepared Seattle Uber movement speed data at the folder datasets/Seattle-movement-data-set
:
hourly_speed_mat_2019_1.npz
(26.4MB)hourly_speed_mat_2019_2.npz
(25.2MB)hourly_speed_mat_2019_3.npz
(31.6MB)
If you want to investigate the missing data problem in Uber movement speed data, please prepare the data in the whole year of 2019 by yourself through the above codes. You can also skip this part and check out our documentation for multivariate time series forecasting on NYC Uber movement speed dataset in the next part.
## Build a speed matrix for the whole year of 2019 in NYC
mat = np.load('hourly_speed_mat_2019_1.npz')['arr_0']
for month in range(2, 13):
mat = np.append(mat, np.load('hourly_speed_mat_2019_{}.npz'.format(month))['arr_0'], axis = 1)
## Calculate missing rates
print('The missing ratte of speed matrix is:')
print(len(np.where(mat == 0)[0]) / (mat.shape[0] * mat.shape[1]))
N, T = mat.shape
sample_rate = np.zeros(T)
for t in range(T):
pos = np.where(mat[:, t] == 0)
sample_rate[t] = len(pos[0]) / N
sample_rate = sample_rate[: 52 * 7 * 24].reshape([52, 24 * 7])
whole_rate = np.mean(sample_rate, axis = 0)
rate = len(np.where(mat == 0)[0]) / (mat.shape[0] * mat.shape[1])
print(rate)
import matplotlib.pyplot as plt
plt.rcParams['font.size'] = 12
fig = plt.figure(figsize = (8, 2))
ax = fig.add_subplot(1, 1, 1)
plt.plot(whole_rate, color = 'red', linewidth = 1.8)
upper = whole_rate + np.std(sample_rate, axis = 0)
lower = whole_rate - np.std(sample_rate, axis = 0)
x_bound = np.append(np.append(np.append(np.array([0, 0]), np.arange(0, 7 * 24)),
np.array([7 * 24 - 1, 7 * 24 - 1])), np.arange(7 * 24 - 1, -1, -1))
y_bound = np.append(np.append(np.append(np.array([upper[0], lower[0]]), lower),
np.array([lower[-1], upper[-1]])), np.flip(upper))
plt.fill(x_bound, y_bound, color = 'red', alpha = 0.2)
plt.axhline(y = rate, color = 'gray', alpha = 0.5, linestyle='dashed')
plt.xticks(np.arange(0, 24 * 7 + 1, 1 * 24))
plt.xlabel('Time (hour)')
plt.ylabel('Missing rate')
plt.grid(axis = 'both', linestyle='dashed', linewidth = 0.1, color = 'gray')
ax.tick_params(direction = "in")
ax.set_xlim([-1, 7 * 24])
# ax.set_ylim([0.6, 1])
plt.show()
# fig.savefig("NYC_missing_rate_stat.pdf", bbox_inches = "tight")
Figure 2: The missing rates of Uber movement speed data aggregated per week over the whole year of 2019. The red curve shows the aggregated missing rates in all 52 weeks. The red area shows the standard deviation of missing rates in each hour over 52 weeks. The 168 time steps refer to 168 hours of Tuesday, Wednesday, Thursday, Friday, Saturday, Sunday, and Monday. (Left panel) The dataset has 98,210 road segments, and the whole missing rate is 64.43%. (Right panel) The dataset has 63,490 road segments, and the whole missing rate is 84.95%.
import numpy as np
mat = np.load('hourly_speed_mat_2019_1.npz')['arr_0']
for month in range(2, 13):
mat = np.append(mat, np.load('hourly_speed_mat_2019_{}.npz'.format(month))['arr_0'], axis = 1)
ratio = np.sum(mat > 0, axis = 1) / (365 * 24)
Print observation rate results:
for threshold in 0.1 * np.arange(1, 10):
print('Observation rate > {0:.2f}'.format(threshold))
print(np.sum(ratio > threshold))
print(np.sum(ratio > threshold) / ratio.shape[0])
print()
In this research, we aim at simultaneously handling the following emerging issues in real-world time series datasets: 1) High-dimensionality (i.e., large
Figure 3: Illustration of the multivariate time series forecasting task on the incomplete data
The temporal matrix factorization consists of matrix factorization (on partially observed time series) and vector autoregressive (VAR) process (on temporal factors).
Figure 4: Illustration of TMF on multivariate time series data. For spatiotemporal traffic states data, it can be factorized into spatial factors and temporal factors in which the temporal factor matrix is indeed a multivariate time series. We aim to impose a certain temporal modeling structure on temporal factor matrix.
We evaluate the proposed model mainly on the Uber movement speed data because these time series data are featured as high-dimensional and sparse. The rolling forecasting scenario is illustrated as follows.
Figure 5: Illustration of the rolling forecasting scenario on multivariate time series data
- Temporal regularized matrix factorization (TRMF)
- Bayesian temporal regularized matrix factorization (BTRMF)
- Temporal matrix factorization (TMF)
- Bayesian temporal matrix factorization (BTMF)
- NoTMF with both seasonal differencing and first-order differencing (NoTMF-1st)
We implement our algorithms with NumPy library, possibly being easy for understanding each line. It should be no barrier for running these codes on any Python environment with a NumPy library. If you have any questions, feel free to create an issue in this repository. If you are looking for an intuitive explanation, please check out our blog post temporal matrix factorization for multivariate time series forecasting on Medium.
Our Python implementation with
numpy
can be adapted to a GPU implementation withcupy
, please replaceimport numpy as np
byimport cupy as np
. It would not difficult to follow the experiment.
- Use MAPE and RMSE as performance metrics.
def compute_mape(var, var_hat):
return np.sum(np.abs(var - var_hat) / var) / var.shape[0]
def compute_rmse(var, var_hat):
return np.sqrt(np.sum((var - var_hat) ** 2) / var.shape[0])
- Generate temporal operators.
def generate_Psi(T, d, season):
Psi = []
for k in range(0, d + 1):
if k == 0:
Psi.append(np.append(np.zeros((T - d - season, d)),
np.append(-1 * np.eye(T - d - season), np.zeros((T - d - season, season)), axis = 1)
+ np.append(np.zeros((T - d - season, season)), np.eye(T - d - season), axis = 1), axis = 1))
else:
Psi.append(np.append(np.append(np.zeros((T - d - season, d - k)),
np.append(-1 * np.eye(T - d - season), np.zeros((T - d - season, season)), axis = 1)
+ np.append(np.zeros((T - d - season, season)), np.eye(T - d - season), axis = 1), axis = 1),
np.zeros((T - d - season, k)), axis = 1))
return Psi
This is a classical approach for defining temporal operators, one effective alternative is defining these temporal operators as sparse arrays in both numpy
and cupy
. Please check out How to define temporal operators as sparse arrays in both numpy
(for CPU) and cupy
(for GPU)?
- Define conjugate gradient for factor matrix.
For an intuitive explanation, please check out using conjugate gradient to solve matrix equations.
def update_cg(var, r, q, Aq, rold):
alpha = rold / np.inner(q, Aq)
var = var + alpha * q
r = r - alpha * Aq
rnew = np.inner(r, r)
q = r + (rnew / rold) * q
return var, r, q, rnew
def ell_w(ind, W, X, rho):
return X @ ((W.T @ X) * ind).T + rho * W
def conj_grad_w(sparse_mat, ind, W, X, rho, maxiter = 5):
rank, dim1 = W.shape
w = np.reshape(W, -1, order = 'F')
r = np.reshape(X @ sparse_mat.T - ell_w(ind, W, X, rho), -1, order = 'F')
q = r.copy()
rold = np.inner(r, r)
for it in range(maxiter):
Q = np.reshape(q, (rank, dim1), order = 'F')
Aq = np.reshape(ell_w(ind, Q, X, rho), -1, order = 'F')
w, r, q, rold = update_cg(w, r, q, Aq, rold)
return np.reshape(w, (rank, dim1), order = 'F')
def ell_x(ind, W, X, A, Psi, d, lmbda, rho):
rank, dim2 = X.shape
temp = np.zeros((d * rank, Psi[0].shape[0]))
for k in range(1, d + 1):
temp[(k - 1) * rank : k * rank, :] = X @ Psi[k].T
temp1 = X @ Psi[0].T - A @ temp
temp2 = np.zeros((rank, dim2))
for k in range(d):
temp2 += A[:, k * rank : (k + 1) * rank].T @ temp1 @ Psi[k + 1]
return W @ ((W.T @ X) * ind) + rho * X + lmbda * (temp1 @ Psi[0] - temp2)
def conj_grad_x(sparse_mat, ind, W, X, A, Psi, d, lmbda, rho, maxiter = 5):
rank, dim2 = X.shape
x = np.reshape(X, -1, order = 'F')
r = np.reshape(W @ sparse_mat - ell_x(ind, W, X, A, Psi, d, lmbda, rho), -1, order = 'F')
q = r.copy()
rold = np.inner(r, r)
for it in range(maxiter):
Q = np.reshape(q, (rank, dim2), order = 'F')
Aq = np.reshape(ell_x(ind, W, Q, A, Psi, d, lmbda, rho), -1, order = 'F')
x, r, q, rold = update_cg(x, r, q, Aq, rold)
return np.reshape(x, (rank, dim2), order = 'F')
- Define nonstationary temporal matrix factorization (
notmf
).
def notmf(dense_mat, sparse_mat, rank, d, lmbda, rho, season, maxiter):
dim1, dim2 = sparse_mat.shape
W = 0.01 * np.random.randn(rank, dim1)
X = 0.01 * np.random.randn(rank, dim2)
A = 0.01 * np.random.randn(rank, d * rank)
if np.isnan(sparse_mat).any() == False:
ind = sparse_mat != 0
pos_test = np.where((dense_mat != 0) & (sparse_mat == 0))
elif np.isnan(sparse_mat).any() == True:
pos_test = np.where((dense_mat != 0) & (np.isnan(sparse_mat)))
ind = ~np.isnan(sparse_mat)
sparse_mat[np.isnan(sparse_mat)] = 0
dense_test = dense_mat[pos_test]
del dense_mat
Psi = generate_Psi(dim2, d, season)
show_iter = 100
temp = np.zeros((d * rank, dim2 - d - season))
for it in range(maxiter):
W = conj_grad_w(sparse_mat, ind, W, X, rho)
X = conj_grad_x(sparse_mat, ind, W, X, A, Psi, d, lmbda, rho)
for k in range(1, d + 1):
temp[(k - 1) * rank : k * rank, :] = X @ Psi[k].T
A = X @ Psi[0].T @ np.linalg.pinv(temp)
mat_hat = W.T @ X
if (it + 1) % show_iter == 0:
temp_hat = mat_hat[pos_test]
print('Iter: {}'.format(it + 1))
print('MAPE: {:.6}'.format(compute_mape(dense_test, temp_hat)))
print('RMSE: {:.6}'.format(compute_rmse(dense_test, temp_hat)))
print()
return mat_hat, W, X, A
def notmf_dic(obs, W, X, A, d, lmbda, rho, season):
dim1, dim2 = obs.shape
rank = X.shape[0]
if np.isnan(obs).any() == False:
ind = obs != 0
elif np.isnan(obs).any() == True:
ind = ~np.isnan(obs)
obs[np.isnan(obs)] = 0
Psi = generate_Psi(dim2, d, season)
X = conj_grad_x(obs, ind, W, X, A, Psi, d, lmbda, rho)
temp = np.zeros((d * rank, dim2 - d - season))
for k in range(1, d + 1):
temp[(k - 1) * rank : k * rank, :] = X @ Psi[k].T
A = X @ Psi[0].T @ np.linalg.pinv(temp)
return X, A
- Define the function for VAR forecasting.
def var4cast(X, A, d, delta, season):
dim1, dim2 = X.shape
X_hat = np.append(X[:, season:] - X[:, : -season], np.zeros((dim1, delta)), axis = 1)
for t in range(delta):
X_hat[:, dim2 - season + t] = A @ X_hat[:, dim2 - season + t - np.arange(1, d + 1)].T.reshape(dim1 * d)
X = np.append(X, np.zeros((dim1, delta)), axis = 1)
for t in range(delta):
X[:, dim2 + t] = X[:, dim2 - season + t] + X_hat[:, dim2 - season + t]
return X
- Define the function for rolling forecasting.
from ipywidgets import IntProgress
from IPython.display import display
def rolling4cast(dense_mat, sparse_mat, pred_step, delta, rank, d, lmbda, rho, season, maxiter):
dim1, T = sparse_mat.shape
start_time = T - pred_step
max_count = int(np.ceil(pred_step / delta))
mat_hat = np.zeros((dim1, max_count * delta))
f = IntProgress(min = 0, max = max_count) # instantiate the bar
display(f) # display the bar
for t in range(max_count):
if t == 0:
_, W, X, A = notmf(dense_mat[:, : start_time], sparse_mat[:, : start_time],
rank, d, lmbda, rho, season, maxiter)
else:
X, A = notmf_dic(sparse_mat[:, : start_time + t * delta], W, X_new, A, d, lmbda, rho, season)
X_new = var4cast(X, A, d, delta, season)
mat_hat[:, t * delta : (t + 1) * delta] = W.T @ X_new[:, - delta :]
f.value = t
small_dense_mat = dense_mat[:, start_time : T]
pos = np.where((small_dense_mat != 0) & (np.invert(np.isnan(small_dense_mat))))
mape = compute_mape(small_dense_mat[pos], mat_hat[pos])
rmse = compute_rmse(small_dense_mat[pos], mat_hat[pos])
print('Prediction MAPE: {:.6}'.format(mape))
print('Prediction RMSE: {:.6}'.format(rmse))
print()
return mat_hat, W, X, A
import numpy as np
dense_mat = np.load('../datasets/NYC-movement-data-set/hourly_speed_mat_2019_1.npz')['arr_0']
for month in range(2, 4):
dense_mat = np.append(dense_mat, np.load('../datasets/NYC-movement-data-set/hourly_speed_mat_2019_{}.npz'.format(month))['arr_0'], axis = 1)
import time
for rank in [10]: # rank of matrix factorization
for delta in [1, 2, 3, 6]: # forecasting time horizon
for d in [6]: # order of VAR
start = time.time()
pred_step = 7 * 24
lmbda = 1
rho = 5
season = 7 * 24
maxiter = 50
mat_hat, W, X, A = rolling4cast(dense_mat[:, : 24 * 7 * 10], dense_mat[:, : 24 * 7 * 10],
pred_step, delta, rank, d, lmbda, rho, season, maxiter)
print('delta = {}'.format(delta))
print('rank R = {}'.format(rank))
print('Order d = {}'.format(d))
end = time.time()
print('Running time: %d seconds'%(end - start))
- Analyzing missing data problem in Uber movement speed data. February 14, 2022.
- Temporal matrix factorization for multivariate time series forecasting. March 20, 2022.
- Forecasting multivariate time series with nonstationary temporal matrix factorization. April 25, 2022.
- Convolution nuclear norm minimization for time series modeling. October 3, 2022.
- Reinforce matrix factorization for time series modeling: Probabilistic sequential matrix factorization. October 5, 2022.
- Discrete convolution and fast Fourier transform explained and implemented step by step. October 19, 2022.
Tracebase provides a collection of time series forecasting approaches and algorithms for you. We aim to support you to explore traffic flow data in urban areas with machine learning models. We provide friendly implementation with a few lines relying on pandas
and numpy
packages.
- Xinyu Chen, Chengyuan Zhang, Xi-Le Zhao, Nicolas Saunier, Lijun Sun (2022). Nonstationary temporal matrix factorization for multivariate time series forecasting. arXiv preprint arXiv:2203.10651. [PDF] [Slides] [Data & Python code]
or
@article{chen2022nonstationary,
title={Nonstationary temporal matrix factorization for multivariate time series forecasting},
author={Chen, Xinyu and Zhang, Chengyuan and Zhao, Xi-Le and Saunier, Nicolas and Sun, Lijun},
journal={arXiv preprint arXiv:2203.10651},
year={2022}
}
Data retrieved from Uber Movement, ยฉ 2022 Uber Technologies, Inc., https://movement.uber.com.
![drawing](graphics/ivado_logo.jpeg)
![drawing](graphics/cirrelt_logo.png)