Title: Two Years of Bayesian Bandits for E-Commerce

Abstract: At Monetate, we've deployed Bayesian bandits (both noncontextual and contextual) to help our clients optimize their e-commerce sites since early 2016. This talk is an overview of the lessons we've learned from both the processes of deploying real-time Bayesian machine learning systems at scale and building a data product on top of these systems that is accessible to non-technical users (marketers). This talk will cover

  • the place of multi-armed bandits in the A/B testing industry,
  • Thompson sampling and the basic theory of Bayesian bandits,
  • Bayesian approaches for accommodating nonstationarity in bandit feedback,
  • user experience challenges in driving adoption of these technologies by nontechnical marketers.

We will focus primarily on noncontextual bandits and give a brief overview of these problems in the contextual setting as time permits.

Bio: Austin Rochford is Principal Data Scientist and Director of Monetate Labs, where he does research and development for machine learning-driven marketing products. He is a recovering mathematician, a passionate Bayesian, and a PyMC3 developer.

A Dockerfile that will produce a container with the dependenceis of this notebook is available here.

In [1]:
%matplotlib inline
In [2]:
from tqdm import tqdm, trange
In [3]:
from matplotlib import pyplot as plt
from matplotlib.dates import DateFormatter, MonthLocator
from matplotlib.ticker import StrMethodFormatter
import numpy as np
import pandas as pd
import scipy as sp
import seaborn as sns
In [4]:
SEED = 76224 # from random.org, for reproducibiliy
np.random.seed(SEED)
In [5]:
sns.set()
C = sns.color_palette()

pct_formatter = StrMethodFormatter('{x:.1%}')

#configure matplotlib
FIGURE_WIDTH = 8
FIGURE_HEIGHT = 6
plt.rc('figure', figsize=(FIGURE_WIDTH, FIGURE_HEIGHT))

LABELSIZE = 14
plt.rc('axes', labelsize=LABELSIZE)
plt.rc('axes', titlesize=LABELSIZE)
plt.rc('figure', titlesize=LABELSIZE)
plt.rc('legend', fontsize=LABELSIZE)
plt.rc('xtick', labelsize=LABELSIZE)
plt.rc('ytick', labelsize=LABELSIZE)

Two Years of Bayesian Bandits for E-Commerce

PyData NYC • October 18, 2018 • @AustinRochford

About Me

Principal Data Scientist and Director of Monetate Labs

@AustinRochford • austinrochford.com • github.com/AustinRochford

[email protected] • [email protected]

About Monetate

  • Founded 2008, web optimization and personalization SaaS
  • Observed 5B impressions and $4.1B in revenue during Cyber Week 2017
monetate.com/about/careers

Non-technical marketer-focused

About this talk

Outline

  • Web optimization
    • A/B testing
    • Multi-armed bandits
  • Bayesian bandits
    • Thompson sampling
  • Bandit bias
    • Inverse propensity weighting

Web Optimization

A/B testing

A/B testing machinery

Ronald Fisher
Whom the Gods Would Destroy, They First Give Real-time Analytics

Sequential testing

Abraham Wald

Sequential optimization

Much of the A/B testing industry uses classical Fisher/Neyman-Pearson style fixed-horizon frequentist significance tests. Sophistocated approaches use sequential hypothesis testing. We've found, through many years of interaction with marketers, that they take a fundamentally Bayesian view of the world. Most interpret p-values as the "probability that the experiment is better/worse than the control."

Multi-armed bandits

Multi-armed bandit problems are extensively studied and come in many variantions. Here we focus on the simplest multi-armed bandit objective, regret minimization.

Multi-armed bandit systems

Bayesian Bandits

Beta-binomial model

$$ \begin{align*} x_A, x_B & = \textrm{number of rewards from users shown variant } A, B \\ x_A & \sim \textrm{Binomial}(n_A, r_A) \\ x_B & \sim \textrm{Binomial}(n_B, r_B) \\ r_A, r_B & \sim \textrm{Beta}(1, 1) \end{align*} $$
$$ \begin{align*} r_A\ |\ n_A, x_A & \sim \textrm{Beta}(x_A + 1, n_A - x_A + 1) \\ r_B\ |\ n_B, x_B & \sim \textrm{Beta}(x_B + 1, n_B - x_B + 1) \end{align*} $$

Thompson sampling

Thompson sampling randomizes user/variant assignment according to the probabilty that each variant maximizes the posterior expected reward.

The probability that a user is assigned variant A is

$$ \begin{align*} P(r_A > r_B\ |\ \mathcal{D}) & = \int_0^1 P(r_A > r\ |\ \mathcal{D})\ \pi_B(r\ |\ \mathcal{D})\ dr \\ & = \int_0^1 \left(\int_r^1 \pi_A(s\ |\ \mathcal{D})\ ds\right)\ \pi_B(r\ |\ \mathcal{D})\ dr \\ & \propto \int_0^1 \left(\int_r^1 s^{\alpha_A - 1} (1 - s)^{\beta_A - 1}\ ds\right) r^{\alpha_B - 1} (1 - r)^{\beta_B - 1}\ dr \end{align*} $$
In [6]:
fig, ax = plt.subplots(
    figsize=(0.75 * FIGURE_WIDTH, 0.75 * FIGURE_WIDTH)
)
ax.set_aspect('equal');

a_post = sp.stats.beta(5 + 1, 15 + 1)
b_post = sp.stats.beta(7 + 1, 13 + 1)

p = np.linspace(0, 1, 100)

px, py = np.meshgrid(p, p)
z = a_post.pdf(px) * b_post.pdf(py) 

ax.contourf(px, py, z, 1000, cmap='BuGn');
ax.plot([0, 1], [0, 1], c='k', ls='--');

ax.xaxis.set_major_formatter(pct_formatter);
ax.set_xlabel(r"$r_A \sim \operatorname{Beta}(6, 16)$");

ax.yaxis.set_major_formatter(pct_formatter);
ax.set_ylabel(r"$r_B \sim \operatorname{Beta}(8, 14)$");

ax.set_title("Posterior reward rates");
In [7]:
fig
Out[7]:
In [8]:
N_SAMPLES = 1000

a_samples = a_post.rvs(size=N_SAMPLES)
b_samples = b_post.rvs(size=N_SAMPLES)

ax.scatter(
    a_samples[b_samples > a_samples],
    b_samples[b_samples > a_samples],
    c='k', alpha=0.25
);
ax.scatter(
    a_samples[b_samples < a_samples],
    b_samples[b_samples < a_samples],
    c=C[2], alpha=0.25
);
In [9]:
ts_fig = fig
In [10]:
fig
Out[10]:
In [11]:
(a_samples > b_samples).mean()
Out[11]:
0.247

Thompson sampling, redeemed

  1. Sample $\hat{r}_A \sim r_A\ |\ n_A, x_A$ and $\hat{r}_B \sim r_B\ |\ n_B, x_B$.
  2. Assign the user variant A if $\hat{r}_A > \hat{r}_B$, otherwise assign them variant B.

Simulating a bandit

In [12]:
class BetaBinomial:
    def __init__(self, a0=1., b0=1.):
        self.a = a0
        self.b = b0
        
    def sample(self):
        return sp.stats.beta.rvs(self.a, self.b)
    
    def update(self, n, x):
        self.a += x
        self.b += n - x
In [13]:
class Bandit:
    def __init__(self, a_post, b_post):
        self.a_post = a_post
        self.b_post = b_post
        
    def assign(self):
        return 1 * (self.a_post.sample() < self.b_post.sample())
    
    def update(self, arm, reward):
        arm_post = self.a_post if arm == 0 else self.b_post
        arm_post.update(1, reward)
In [14]:
def generate_rewards(a_rate, b_rate, n):
    yield from zip(
            np.random.binomial(1, a_rate, size=n),
            np.random.binomial(1, b_rate, size=n)
        )
In [15]:
A_RATE, B_RATE = 0.05, 0.1
N = 1000

rewards_gen = generate_rewards(A_RATE, B_RATE, N)
In [16]:
bandit = Bandit(BetaBinomial(), BetaBinomial())
arms = np.empty(N, dtype=np.int64)
rewards = np.empty(N)

for t, arm_rewards in tqdm(enumerate(rewards_gen), total=N):
    arms[t] = bandit.assign()
    rewards[t] = arm_rewards[arms[t]]

    bandit.update(arms[t], rewards[t])

100%|██████████| 1000/1000 [00:00<00:00, 5443.70it/s]
In [17]:
fig, (regret_ax, assign_ax) = plt.subplots(
    ncols=2, sharex=True, 
    figsize=(2 * FIGURE_WIDTH, 6)
)

plot_t = np.arange(N) + 1

regret_ax.plot(
    plot_t, rewards.cumsum(),
    c=C[0], label="Bandit"
);

regret_ax.plot(
    plot_t, A_RATE * plot_t,
    c=C[1], ls='--',
    label="Variant A"
);

regret_ax.plot(
    plot_t, B_RATE * plot_t,
    c=C[2], ls='--',
    label="Variant B"
);

regret_ax.plot(
    plot_t,
    0.5 * (A_RATE + B_RATE) * plot_t,
    c='k', ls='--',
    label="A/B testing"
);

regret_ax.set_xlim(1, N);
regret_ax.set_xlabel("Sessions");

regret_ax.set_ylabel("Rewards");

regret_ax.legend(loc=2);

assign_ax.plot(plot_t, (arms == 1).cumsum() / plot_t);

assign_ax.hlines(
    0.5, 1, N,
    colors='k', linestyles='--'
);

assign_ax.set_xlabel("Sessions");

assign_ax.set_ylim(0, 1);
assign_ax.yaxis.set_major_formatter(pct_formatter);
assign_ax.set_yticks(np.linspace(0, 1, 5));
assign_ax.set_ylabel("Traffic shown variant B");

fig.suptitle("Cumulative bandit performance");
In [18]:
fig
Out[18]:

Simulating many bandits

In [19]:
def simulate_bandit(rewards, n,
                    progressbar=False,
                    prior=BetaBinomial):
    bandit = Bandit(prior(), prior())

    arms = np.empty(n, dtype=np.int64)
    obs_rewards = np.empty(n)
    
    if progressbar:
        rewards = tqdm(rewards, total=n)

    for t, arm_rewards in enumerate(rewards):
        arms[t] = bandit.assign()
        obs_rewards[t] = arm_rewards[arms[t]]

        bandit.update(arms[t], obs_rewards[t])
        
    return arms, obs_rewards
In [20]:
N_BANDIT = 100

arms = np.empty((N_BANDIT, N), dtype=np.int64)
rewards = np.empty((N_BANDIT, N))

for i in trange(N_BANDIT):
    arms[i], rewards[i] = simulate_bandit(
        generate_rewards(A_RATE, B_RATE, N), N
    )

100%|██████████| 100/100 [00:11<00:00,  8.98it/s]
In [21]:
fig, (regret_ax, assign_ax) = plt.subplots(
    ncols=2, sharex=True, 
    figsize=(2 * FIGURE_WIDTH, 6)
)

plot_t = np.arange(N) + 1

cum_rewards = rewards.cumsum(axis=1)

ALPHA = 0.1
low_rewards, high_rewards = np.percentile(
    cum_rewards, [100 * ALPHA / 2, 100 * (1 - ALPHA / 2)],
    axis=0
)

regret_ax.fill_between(
    plot_t, low_rewards, high_rewards,
    color=C[0], alpha=0.5,
    label=f"{1 - ALPHA:.0%} interval"
);
regret_ax.plot(
    plot_t, cum_rewards.mean(axis=0),
    c=C[0], label="Bandit (average)"
);

regret_ax.plot(
    plot_t, A_RATE * plot_t,
    c=C[1], ls='--',
    label="Variant A"
);

regret_ax.plot(
    plot_t, B_RATE * plot_t,
    c=C[2], ls='--',
    label="Variant B"
);

regret_ax.plot(
    plot_t,
    0.5 * (A_RATE + B_RATE) * plot_t,
    c='k', ls='--',
    label="A/B testing"
);

regret_ax.set_xlim(1, N);
regret_ax.set_xlabel("Sessions");

regret_ax.set_ylabel("Rewards");

regret_ax.legend(loc=2);

cum_winner = (arms == 1 * (A_RATE < B_RATE)).cumsum(axis=1)
low_winner, high_winner = np.percentile(
    cum_winner / plot_t, [100 * ALPHA / 2, 100 * (1 - ALPHA / 2)],
    axis=0
)

assign_ax.fill_between(
    plot_t, low_winner, high_winner,
    color=C[0], alpha=0.5,
    label=f"{1 - ALPHA:.0%} interval"
);

assign_ax.plot(plot_t, cum_winner.mean(axis=0) / plot_t);

assign_ax.hlines(
    0.5, 1, N,
    colors='k', linestyles='--'
);

assign_ax.set_xlabel("Sessions");

assign_ax.set_ylim(0, 1);
assign_ax.yaxis.set_major_formatter(pct_formatter);
assign_ax.set_yticks(np.linspace(0, 1, 5));
assign_ax.set_ylabel("Traffic shown better variant");

fig.suptitle(f"Cumulative bandit performance ({N_BANDIT} simulations)");