A payment card (such as debit or credit) is one of the most convenient payment methods for purchasing goods and services. Hundreds of millions of card transactions take place across the globe every day, generating a massive volume of transaction data. The data render a holistic view of cardholder-merchant interactions, containing insights that can benefit various applications, such as payment fraud detection and merchant recommendation. However, utilizing these insights often requires additional information about merchants missing from the data owner's (i.e., payment company's) perspective. For example, payment companies do not know the exact type of product a merchant serves. Collecting merchant attributes from external sources for commercial purposes can be expensive. Motivated by this limitation, we aim to infer latent merchant attributes from transaction data. As proof of concept, we concentrate on restaurants and infer the cuisine types of restaurants from transactions. To this end, we present a framework for inferring the cuisine types of restaurants from transaction data. Our proposed framework consists of three steps. In the first step, we generate cuisine labels for a limited number of restaurants via weak supervision. In the second step, we extract a wide variety of statistical features and neural embeddings from the restaurant transactions. In the third step, we use deep neural networks (DNNs) to infer the remaining restaurants' cuisine types. The proposed framework achieved a 76.2% accuracy in classifying the US restaurants. To the best of our knowledge, this is the first framework to infer the cuisine types of restaurants by analyzing transaction data as the only source.

1. Profiling US Restaurants from Billions
of Payment Card Transactions
Himel Dev Hossein Hamooni

2. Payment Card Transactions
Across the globe, billions of
people regularly use payment
cards (say debit and credit) to
pay for goods and services.
Payment processing companies
handle these transactions and
record their attributes in the
form of transaction data.
Transaction data contain rich
insights into the behavior of
cardholders and the business
of merchants.
Transactional insights can benefit various applications, such
as payment fraud detection and merchant recommendation.

3. However, utilizing transactional insights often require
auxiliary information about merchants that are missing
from the payment company’s perspective.

4. Can’t We Just Use Google orYelp?
Cost: It is expensive to acquire
merchant information, especially for
commercial purposes.
Unavailability: Payment cards are
used in many countries where crowd
sourced merchant data is unavailable.
Unreliability: Acquired information
become outdated, as new merchants
appear or old merchants disappear.

5. Our big goal is to infer latent merchant attributes from
transaction data, without using external sources.

6. Proof-of-Concept Use Case
Infer the cuisine types of restaurants by analyzing only transaction data.
Restaurant
Recommendation
Fraud
Detection

7. Transaction Data
Our dataset contains four billion debit and credit card transactions in
more than half a million US restaurants within three months.
We aim to develop a framework for inferring the cuisine types of
restaurants from the transaction data.
Cardholder Merchant Zip Date & Time $$ $$+
5r8g3d0u5q Peking Cafe 55075 20-06-2017 18:03:05 12.65 13.65
5r8g3d0u5q Health Junkie 55075 12-07-2017 19:21:17 17.81 19.81
5r8g3d0u5q Pineda Tacos 55075 13-07-2017 13:06:04 10.99 12.00

8. 1
Cardholder Merchant Zip Date & Time $$ $$+
5r8g3d0u5q Peking Cafe 55075 20-06-2017 18:03:05 12.65 13.65
5r8g3d0u5q Health Junkie 55075 12-07-2017 19:21:17 17.81 19.81
5r8g3d0u5q Pineda Tacos 55075 13-07-2017 13:06:04 10.99 12.00
Location PriceTime
Group CapacityTips
Loyalty
Name Label
Peking
Cafe
Chinese
Pineda
Tacos
Mexican
Health
Junkie
? Embedding
2
2
3
Deep Neural Network
Cuisine Inference
Framework
Weakly Supervised
Label Generation
Statistical Feature and
Neural Embedding Extraction
Deep Neural Network
Based Classification
1
2
3

9. 1.Weakly Supervised Label Generation

10. Cuisine Taxonomy Creation
Id Cuisine Type Subcategories
1 Latin American Mexican, Cuban, Brazilian, Colombian
2 European French, Italian, German, Polish, Irish
3 Mediterranean Greek,Turkish
Middle Eastern Saudi Arabian, Lebanese, Persian,Afghan
African Moroccan, Ethiopian, Eritrean
4 South Asian Indian, Pakistani, Nepalese, Bangladeshi
5 South East Asian Thai,Vietnamese, Indonesian, Malaysian
6 East Asian Chinese, Japanese, Korean, Mongolian
7 Grill and Steak Grill, Steakhouse
8 Fastfood Sandwich, Burger, Pizza
9 Bar Bar, Pub,Tavern, Inn
10 Dessert Ice Cream, Cafe, Bakery, Juice
We create a cuisine
taxonomy for the US
restaurants.
Our taxonomy contains
the ten most popular
cuisine types in the US.
Each of these major
cuisine types cover many
minor cuisine types.

11. Seed Word Compilation
Restaurant Name Cuisine Type
Peking Garden Chinese
Golden Wok Chinese
Ambar India Indian
Himalayan Chimney Indian
Biaggi's Ristorante Italiano Italian
Pizzeria Antica Italian
Maize Mexican Grill Mexican
Burrito King Mexican
Garbanzo Mediterranean Fresh Mediterranean
Jerusalem Mediterranean
Good Fella ???
We compile a set of seed words
for each cuisine type in our
taxonomy.
We use these words as common
patterns to generate cuisine labels
for restaurant names.
Currently, we have a list of 225
seed words that represent the ten
major cuisine types.

12. Bootstrapped Label Expansion
We extract new (beyond seed) words from restaurant names to utilize as highly
accurate patterns for increasing the coverage of labeled restaurants.
Frequency: The word needs to appear in θf fraction of all restaurant
names
Precision: If we use the word and its majority label as a labeling rule, the
rule needs to be true for θp fraction of labeled restaurants
Significance: The ratio of labeled and unlabeled restaurants that contain the
word should be θs
Using seed and bootstrapped words, we
could label 35% restaurants in our dataset.

13. Topic Modeling
To augment the keyword-based approach, we develop a custom topic model.
Issue Description Solution
Monolith Many restaurant names consist of a single word Sprinkling
Sparse Sparse word co-occurrence patterns in restaurant names BTM
LongTail Long-tail distribution of words in restaurant names Stratification
The resultant topics (cuisine types) are coherent and consistent
with the cuisine types generated by our keyword-based approach.

14. 1I. Statistical Feature and Neural Embedding Extraction

15. Statistical Features
Feature Type Description
Pricing The deciles of authorized amount in transactions
Tipping culture The deciles of (settlement amount - authorized amount) in transactions
Serving capacity The deciles of hourly transaction count
Party size The proportion of transactions for different party size
Party pricing The average authorized amount for different party size
Temporal pattern I The distribution of number of transactions over days of the week
Temporal pattern II The distribution of number of transactions over the hours of weekdays
Temporal pattern III The distribution of number of transactions over the hours of weekends
Customer revisitation The deciles of the number of revisits by the customers
Customer loyalty The deciles of the number of restaurants visited by the customers
Location The digits of restaurant zipcode and corresponding location granularity

16. Statistical Feature Insights
X

17. Customer Restaurant Interaction
U1 U2 U3 U4
R1 R2 R3 R4 R5

18. Micro and Macro Hypotheses
The distinction between the hypotheses lies in application: individual vs group.
Micro Hypothesis: The compatibility between a customer’s preferences and a
restaurant’s attributes is a good predictor of whether the customer will visit
the restaurant. For example, a vegetarian is likely to visit an Indian restaurant.
Macro Hypothesis: The type of customers who visit a given restaurant (as a
whole) is a good predictor of its attributes. For example, a restaurant is
unlikely to be a steakhouse if many of its customers are vegetarian.

19. Micro Embedding
U1 U2 U3 U4
R1 R2 R3 R4 R5
R1 R3R2 R2 R5 R4Micro: word2vec

20. Macro Embedding
U1 U2 U3 U4
R1 R2 R3 R4 R5
U1 U3U2 U2 U4 U3Macro: doc2vec

21. Micro and Macro Embedding
U1 U2 U3 U4
R1 R2 R3 R4 R5
U1 U3U2 U2 U4 U3
R1 R3R2 R2 R5 R4Micro: word2vec
Macro: doc2vec

22. Name Embedding
We generate name embeddings to utilize the non-labeling words in names.
We first remove the labeling words from each restaurant name.
We then retrieve pre-trained GloVe embedding for each remaining word in name.
We finally combine the word embeddings via max pooling.
We generate three sets of restaurant embeddings
to represent the latent characteristics of restaurants.

23. 1II. Deep Neural Network Based Classification

24. DNN Models
Shallow Feedforward: This is a feedforward neural network with two hidden layers
Deep Feedforward: This is a feedforward neural network with four hidden layers
Deep Feedforward Res: This is a deep feedforward network with residual connections
Wide and Deep: This is the wide and deep network that captures feature interaction
Deep and Cross: This is the deep and cross network that applies feature crossing
To demonstrate the effectiveness of our framework, we develop several DNN models.

25. Price Tips Location…
Statistical Features
Micro Macro Name
Embeddings
Concatenated Layer
Hidden Layer 1
Hidden Layer 2
S
Output
…
…
+ + + + + +
Deep
Feedforward
with Residual

26. Experimental Evaluation

27. Performance Comparison
The Deep Feedforward Network outperforms Wide and Deep, and Deep and Cross.
Adding residual connections boost the performance of the Deep Feedforward Network.
DNN Model Accuracy
Shallow Feedforward 0.743
Deep Feedforward 0.756
Deep Feedforward with Residual 0.762
Wide and Deep 0.740
Deep and Cross 0.746

28. Confusion Matrix

29. Confusion Matrix

30. Confusion Matrix

31. Ablation Study

32. Ablation Study

33. Ablation Study

34. Summary
We developed a framework for inferring the cuisine types of restaurants from debit
and credit card transactions.
Our proposed framework consists of three steps: 1) weakly-supervised label
generation, 2) statistical feature and neural embedding extraction, and 3) deep neural
network based classification.
The proposed framework achieved a 76.2% accuracy in classifying the US restaurants.