E-Commerce companies face a wide variety of fraudulent activities. Machine Learning models are continuously built to detect them for their efective mitigation. Randomly typed alphanumeric characters as customer addresses is one peculiar fraud. The orders are placed with such addresses. We refer to them as monkey-typed addresses. Accurate methods of identifying such addresses is important in reducing operational cost. The current work presents machine learning based approaches that classify a given address as normal or monkey-typed with a high accuracy. The approach integrates address preprocessing, novel feature generation and classification.
Online e-commerce companies facilitate its users to order items
online and deliver them to the suggested addresses. Users register
themselves by entering one or more of their addresses that would
ensure that an ordered shipment is positively delivered to them.
Thus the customer addresses form important part of e-commerce
companies. A number of challenges in relation to the addresses,
especially, in developing countries include[
In order to avail price discounts or with possible fraudulent intent, there are occasions where orders are placed with addresses containing randomly typed alphanumeric characters. We refer to them as monkey-typed addresses in this work. Early identification of such addresses in the operations chain is necessary to reduce operations cost. For example, when such addresses remain undetected, it would lead to activation of a number of stages in the operations
E-Commerce companies face a wide variety of fraudulent
activities. They include payment frauds, return frauds, delivery frauds,
lost-and-stolen articles, fake reviews and fraudulent ratings. Such
frauds are prevalent in both developing and developed countries.
They are well discussed and many solutions are proposed [
We have not come across any peer-reviewed, published work on
address fraud. Most of the discussions in blogs that relate address
fraud [
E-commerce companies face a number of fraudulent activities. They include frauds caused by buyers, sellers and logistic partners. Some relate to exploiting discounts ofered by the companies to resell the products for profit. The activities also include number of fraudulent claims and counter claims about non-receipt of a genuine product, claims of receipt of part of shipments as against a promised list, etc. Some are more subtle and dificult to detect. We discuss some such frauds and their solutions.
Credit card fraud detection is one important activity. They are
broadly classified into two [
We propose two approaches to classify a given user address as anomalous or genuine. We discuss them in detail in Section 3.
The problem can be defined as, classification of a given address into either normal or monkey-typed accurately and eficiently. Eficiency is required for real time detection of such fraud in customer on-boarding process. The solution is also useful to clean the historical database of customer data. A study of monkey-typed and normal addresses initially indicated that a possible solution can be obtained through language modeling. We experimented with standard language models and found as not efective as shown later in Section 3.3. In view of this, we propose to solve this problem in the following two ways of pattern representation.
(1) Approach-1: In this approach, we derive text features as
groups of characters. Through machine learning models, we
learn from data of those combinations of characters that are
normal or abnormal and culminate in classification of an
address.
(2) Approach-2: In this approach, we exploit importance of
vowels in English [
We achieve this objective through the following steps of the proposed methods.
(1) Preprocess the addresses: The addresses at the time of
entry would contain a number of special characters such
as $, ^,%, *,/, etc. Earlier studies [
For classifying a given address as normal or monkey-typed, it is not possible to obtain all combinations of monkey-typing possibilities. We initially place our eforts in understanding the conventional addresses as well as monkey-typed addresses. Subsequently we devise methods for efective classification of such addresses. We discuss about the nuances of the data and its transformation as it passes through important stages of preprocessing. 3.1
The number of addresses is in the order of millions. Monkey-typed addresses form a very small percentage of the total address set. Table 2 contains a sample of normal addresses. The table contains addresses as entered by the registered users. In the table, in order to maintain privacy of the users, we replaced specific dwelling identity with x’s. It can be noticed from the data that a number of challenges exist in processing these addresses. The details that could have been provided in subsequent stages of address entry, such as city name, and ZIP code would form part of the main address. It should further be noted that some city names are inadvertently spelled in diferent ways. Some examples are BANAGALORE in place of BANGALORE and Bhubaneswar in place of Bhubaneshwar. The preprocessing models also need to account for some city names that are renamed oficially such as { Benguluru, Bangalore} or {Gurgaon, Gurgram}. We Spell variants of a word apartment as entered by users apartment apartmenet apartmanet apartmenst apartmenrt apartmennt apartment appratment aparatment appratmant aparment apartent apprment apartement apartament apartnment apprtement apartmment aparttment apartrment apartememt apaprtment aparrtment apaertment apartmt apartme apartmernt apartmeant apartmemt apartmebt apatrtments apaternments apasrtments apartmensts apartmenmts aparntments observe presence of many variants of these words when we study a large set of customer entered addresses.
For example, in a dataset of about 10000 addresses of Bangalore,
we notice about 18 diferent spell variations of an area name
“koramangala", 61 variants of “Bannerghatta Road", 263 variants of the
word "apartments". The variants are obtained through clustering
the addresses [
Table 3 contains monkey-typed, and potentially fraudulent addresses. It can be noted that they contain many combinations of characters. Interestingly, some intelligent fraud customers, leave spaces while entering monkey typed addresses, to make them appear as normal addresses which typically contain multiple words. The study of such addresses provide interesting insights. It is noticed that such customers also write random characters with a mix of upper and lower case characters. Thus the proposed solution should overcome the issues of inter-word spaces, and mixed case letters. Further, in such randomly typed letters, either vowels are extremely rare or completely dominate and remain repetitive.
Figure 1 contains a word-cloud of a sample of such addresses after splitting them into 4-character long strings. We can observe from the figure that frequently used substrings predominantly match with successive letters on qwerty keyboard although there exist other combinations too.
We will further discuss challenges in the data and preprocessing methods that are necessary for working with the proposed approach in the following section. 3.2
Both the classes of addresses of normal and monkey-typed have many specific observations. Normal address data have the following challenges.
(1) Control and special characters, that are inadvertently introduced during data entry, storage and retrieval. (2) Upper and lower case letters (3) Many spell variants of a given word (4) Missing spaces, such as, IndustrialLayout in place of Industrial Layout Sl. No.
1 2 3 4 5
(5) Undesirable spaces such as, Industrial Lay out in place of
Industrial Layout
(6) Terms that do not help classification[
The monkey-typed addresses data have the following challenges. (1) Variable length strings with maximum number reaching nearly 100 characters. (2) Many of them have repeated substrings such as asdfasdfgij etc. (3) Some addresses are provided as multiple monkey-typed strings so as to mimic a normal address. (4) Combination of upper and lower case characters
The above discussion presents a list of all the challenges present in addresses. For the problem under consideration, some challenges like accidental merging or separation of words is not important since we ultimately remove the spaces in the address while generating features as we present in the following discussion.
We take following preprocessing or data preparation steps while generating features.
(1) Removing control and special characters, such as, $, ^,%, *,/,
(2) Reducing all data to lower case
(3) Generating equivalent set of words with the help of
clustering of empirical data [
Table 4 illustrates processing of the addresses for Approach-1. It contains addresses in three stages such as (a) raw form as entered by customers, (b) preprocessed and (c) feature form. We provide illustrations for both 8-character and 4-character long features. Different feature subsequence lengths are studied through experiments and reported in Section 4. In the table, we consider two normally written addresses and three monkey-typed addresses. It is interesting to note the variety of letter combinations and word spaces that fraudsters use in writing monkey-typed address. The ‘*’s at the end of a word indicate blanks that are necessary to complete 8character bound. It should however be noted here that as discussed in Section 3.2, repeat substrings are removed.
Table 5 contains features for the preprocessed addresses for
Approach-2. The preprocessed addresses are taken from Table 4.
We consider each of these binary digits as features.
Initial approach to solve this problem is by using statistical
language models, where a probability is assigned to a character
sequence, {w1, w2, . . . , wn }. It is expected that the scores for monkey
typed addresses such as
gtysfkjhjuhkjkeraladjuhgjdhiidjidjidkgj will be distinctly diferent from that of a normal address. We
conducted experiments with kylm language modeling toolkit [
Monkey typed addresses occur very rarely in a given dataset. But they are present and need to be located through a model. The ratio of normal addresses to monkey typed addresses leads to severe class imbalance. Here we consider equal number of patterns for each of the classes to demonstrate working of our proposed methods. We consider 30,000 addresses with nearly equal number of normal and monkey-typed addresses. We present case studies and their corresponding datasets in Table 7. For cases 1-3, we consider 23 : 13 division between training and test datasets. As a first step, the data is randomly shufled. Each time an exclusive set of training dataset of size 23 , is chosen and the remaining data is considered as test data, thus resulting in three distinct cases. For cases 4-5, the training and test data are considered as 12 : 12 .
Table 7 contains serial number of case studies and corresponding datasets. The pattern numbers presented in Column 2 and 3 of the table correspond to training and test datasets of each of the classes. For example, Case-3 indicates that from randomly shufled training data corresponding to class-1 and class-2, {1 to 5000} and {10001 to 15000} patterns each are considered for training and the rest as test data. The training data and their corresponding labels are considered accordingly.
The data is subjected to 2-class classification using linear SVM and Naive Bayes classifier. The performance is comparable. For the sake of brevity, we present the results of the exercises using linear SVM alone for both the methods. Table 8 contains results corresponding to Approach-1. The precision, recall and F-score numbers are presented in columns 2,3 and 4 for the cases considered in Table 7.
Experiments are carried out with diferent character lengths for the features such as 4,8, and 16. We observe that the performance is better for 4 and 8. Further, the results with 4-character long features show better performance as evident from the table.
Performance of 4-character long features perform nearly the same for both the data splits of 23 : 13 and 12 : 12 . The best and average classification accuracy of 98% is obtained for features of length 4 and training-test data split of 12 : 12 .
Table 9 contains results with Approach-2 for the same cases as listed in Table 7. The performance of all the five cases is nearly the same. Approach-1 shows superior performance to Approach-2, in general.
The true positive rate and false negative rates for Case-4 of Table 7 for diferent proportion of monkey-typed and normal addresses are provided in Table 10. This is provided in order to indicate performance under varying natural rates of monkey typed addresses ranging from 5% to 75%.
Although Approach-1 outperforms the other approach, deploying the solution across the country requires training of diferent combination of k-char groups, since the language spoken, names of locality, and ethnic words are distinct in diferent regions. Approach2 is immune to basic nuances of words and focus on pronounceability of words. Thus, it is a more generic pattern representation.
The classification time per pattern is 38 microsec on Intel i3, 2.6Ghz, 4-core, 8GB RAM machine. This is well within the single digit millisec latency requirement of the operational system.
It is interesting to see how some failures have happened in the classification. We look at an example of a normal address that is classified as monkey typed. The address ’microwave lab Dept: ECEE, C04, IISC’, when split into 4-char features, amounts to ’micr owav elab dept ecee c04i isc*’. Such combination of characters, such as ‘elab’, ‘isc*’, ’owav’, etc., occur in monkey typed addresses as well. This is the possible reason for its misclassification. Consider the address which is monkey typed but classified as normal address, viz., ’figb irds igjr educ tion sjfk kuit ebgo b*** . It has mix of monkey-typed 4-char features, such as, ‘igjr’,‘sjfk’, and ‘edgo’ and few features that occur in normal addresses, such as, ‘educ’, and ‘tion’. This possibly resulted in its classification as a normal address.
An important aspect of the deployment is dealing with class imbalance. We provide a discussion in Section 5 on various approaches to tackle imbalance.
Like in many fraud or spam classification problems, monkey-typed
addresses form a very small percentage of combined class dataset
leading to severe class imbalance [
Consider a 2-class classification scenario, where the number
of patterns belonging to one class far exceeds those belonging to
second class. Such data is termed imbalanced between the classes.
Finding an efective classification model in such a scenario is
challenging, since many classification models favour the majority class.
In many practical problems, the minority class represents
fraudulent patterns. Misclassification of fraudulent patterns as genuine
ones is expensive. Some conventional approaches to solve the
problem are under-sampling majority class till balance is obtained,
oversampling the minority class in similar manner and choice of unequal
classification costs. In case of under-sampling, when
representative patterns intra-class variations are eliminated, the classification
accuracy sufers. In case of over-sampling of existing patterns of
minority class, it is not guaranteed to arrive at inter-class
discriminative boundary. In the last 15 years, a number of intelligent sampling
methods, boosting and combination of sampling and boosting, are
studied to efectively overcome these issues [
Adapting the above discussed approaches would help mitigating class imbalance problem for monkey-typed address classification.
We consider a practical problem that afects e-commerce companies where potentially fraudulent addresses containing a single or multiple strings of randomly typed addresses occur. We refer to them as monkey-typed addresses. When such addresses remain undetected, they would amount to loss of revenue and cause strain on resources at various stages of processing, starting from order checkout to logistics. The problem is challenging. Through an initial study, we observed that conventional language modeling approaches do not generalise well. We present two novel solutions (a) by forming ifxed-length features and (b) by exploiting inter-vowel distance, after subjecting the addresses to a number of preprocessing steps. We present the data challenges, preprocessing steps, proposed approaches and feature generation. We carry out experiments for feature lengths of 4 and 8 for Approach-1 and features marking vowels for Approach-2. The performance is similar in both the approaches.
Further work includes reducing the number of features through optimal feature selection approaches, experimenting with models that could work across diferent geographical regions within a large country and adapting to class imbalance.
The authors would like to thank Aditya Rachakonda for his helpful comments.