In 2015, the number of digital buyers was 1.46 billion. In 2021, the figure has grown to 2.14 billion. Such a growing trend inspires businesses to look for innovative solutions to the client’s pains and integrate AI/ML tools into their ecommerce business operations.
Product matching in e-commerce has recently become a frequently discussed topic among retailers. They use product matching to ensure a flawless buying experience and take advantage of selling online.
Some retailers provide unique offers to their clients, while others offer identical products and put effort to stay competitive. In this article, we will talk about product matching in ecommerce and explore deep learning as a tool for identifying the same offers.
What Is Product Matching?
The Internet and mobile technologies have changed the way people do shopping. Both retailers and buyers are entering the era of advanced ecommerce where purchases, transactions, and order fulfilment are made quickly and effortlessly. As a part of the strategy for facilitating buying experience, ecommerce businesses implement product matching.
Nowadays, people look through various e-commerce platforms in an effort to buy the desired product at the best price. The most popular ecommerce platforms are Shopify, Magento, Squarespace, WooCommerce, etc. Magento, for example, powers nearly 12% of all ecommerce websites globally, which points at the platform’s convenience for both retailers and consumers.
Between 2017 and 2018, the number of Magento websites almost doubled. Today, this figure keeps growing, and we now have more than 250 000 active Magento websites worldwide. If you are going to build your eCommerce store and need help with Magento development, contact the Forbytes team and we will gladly help.
Dwelling upon the topic, retailers may use different ecommerce spaces for selling their products. That is, it happens that identical products are offered by different retailers on the same platform. A user searches for a specific product and sees the same offer made by different sellers in one place. The consumer compares product price, attributes, quality and then chooses the best offer.
The situation is analogous to the case when a buyer visits multiple physical stores in the search of the most advantageous offer. The difference is that an online decision is made effortlessly and quickly, without the need to leave home. What makes it possible for a buyer to make an informed and advantageous choice based on product comparison? The answer is product matching.
How Sellers and Byers Benefit from Product Matching in eCommerce
Product matching in ecommerce denotes matching the same products offered by different sellers in one search result with the help of deep learning. Product matching is important for both retailers and online shoppers.
For consumers, product matching creates the opportunity to choose the best offer after comparing all available options. Suppose that a consumer wants to buy a table lamp. The main argument in favor of a specific offer will be a price. They go to a particular marketplace and start comparing the options.
It turns out that two sellers offer the same table lamp for $50. Meantime, another seller says “buy the lamp for the $53 and get a table clock for free.” The price-quality ratio seems more attractive in the second case. If a client needs a table clock, the second option will be more valuable for them. To enable comparison, product matching puts the same offers into one search result and ensures flexibility and convenience for a client.
For a seller, product matching can be used for developing a rational price policy and keeping a business competitive. Comparing product prices across the platform allows businesses to learn more about their competitors and make their own offers stand out from the crowd. The special promotion where the table lamp is sold with the table clock is an example of how a seller can differentiate.
Besides, product matching in ecommerce becomes useful when a retailer wants to formulate their offer properly. To make clients find their product, businesses have to name it in a particular way, use specific images, and describe product attributes precisely. Product matching helps to outline tendencies and learn from the competitors’ behavior.
Product Matching Models
Usually, products offered on different platforms are described by a title, attributes, and image. A product title is usually a brief text identifying the key information about a product. It can be a product name and its characteristics. For instance, “plaid shirt male.” Product attributes provide more details on the product and are usually based on name-value pairing.
To unify the way sellers describe product attributes, categorization or structured tables are used. If we talk about the male plaid shirt, product attributes described on a website may include color and size options, the information about fabric and the manufacturer. Images illustrate how a product looks like, and many sellers borrow the same image from one another.
When an image is the same, it is easy to match identical options. Yet, it also reduces the chances that a client will pay attention to your offer in the background of competitors. But how to enable product matching and do not allow technology to make a mistake while matching identical options? This is what we will discuss below.
The title similarity approach to product matching in ecommerce allows ML to compare the same offers by quantifying the similarity of the titles. It enables ML to easily detect contextually same titles even when the strings under comparison significantly differ.
Let’s take an example of several sellers making the same offer of an iPhone XS Max. This is what they put in the title:
- iPhone XS Max
- iPhone XS Max 6.46-inch
- Apple iPhone XS Max
- iPhone XS Max Black
- iPhone XS Max 256 GB NEW
As you may see from the given list, the titles for the same product differ from seller to seller. In this case, to detect the same offers, ML should use a particular product matching algorithm illustrated below. The algorithm will help to evaluate the degree of similarity and put the same products in the same search result.
The first step in the algorithm is preprocessing based on pointwise mutual information. Preprocessing allows ML to see two different tokens as a single entity. This, in turn, enables us to further compute the word-level embeddings. In the first layer, word-level embeddings are trained on the title data from the whole catalog. Training enables the deep learning technology to more effectively and correctly handle data that has not been trained initially.
Next, the concatenated padded titles are used to train a convolutional network and ensure that title length is equal in every option. For this purpose, the skip-gram model can be utilized. And what if one title contains more/less word-level information than another one? To prevent errors, we take a random title and pair it with the same title that randomly lacks some tokens. After adding them as a matched pair, we expand the capability of the title similarity measure.
Product similarity is identified based on the comparison of prices. For instance, there is a range of the same products with approximately the same pricing. One offer stands out from the rest, which may point to the fact that the product is different. This principle works vice versa: to detect similar offers, price distribution is analyzed. In case of similarity, products are displayed in the same range.
There are two common ways to detect price similarities. The first, price outlier detection, is used to detect price similarity when one price is compared to a group of products with similar prices. The test is used in cases when a price is higher or lower compared to the pricing of the whole product group. The second, clustering, helps understand the volume of similar products based on their price. The learned data can be applied as a feature in the overarching system of product matching algorithms.
Under this model, product matching is based on analyzing product categories. These can be item size, brand, color, condition, model. With this product matching algorithm, the technology analyzes structured data and measures the level of discrepancy between products. In case of low discrepancy, the items are considered the same. The following network can be used for extracting the product attributes:
Product attributes can fall into 2 categories: limited range values and endless values. Limited range values have a fixed range of values. With the help of one-hot encoded vectors, these values can be transformed into ML-processed formats. Also, we can apply a convolutional neural network, similar to the one in the title similarity algorithm. It will differ in the output (since there will be only one product title) and in the last layer (since it will be changed to SoftMax).
Meantime, endless values do not have a fixed range. As the range is growing, the accuracy of ML models takes a hit. To solve this problem, we apply attribute extraction in the form of sequence labeling. Under this scheme, every title is tokenized and gets one of the three labels. The labels denote the first or intermediate token or mark that the token is not a part of a brand name.
Image similarity is based on the similar principle as title similarity. To detect the same products, the process of quantifying image similarity is involved. The main challenge of image similarity occurs where there is not enough labeled data on the product. Moreover, the images of the same product may differ in perspective (as seen below), color temperature, color brightness, etc.
To reduce the volume of manual processing of image pairs, you can apply auxiliary taxonomy to training image-based models. This indirect method allows us to view a particular product in a chain of nodes in the taxonomy. For example, a product may belong to the Clothing > Woman > Blazers node.
Under the given product matching algorithm, we use the first layers to extract features and compute cosine similarity. It is better to apply several architectures. The more models you use, the more idiosyncrasies you will be able to handle in the future.
How Else Can Product Matching in eCommerce Be Used
Apart from helping a client find the right product, ML-powered matching algorithms are useful in the following cases:
Making Relevant Product Recommendations
With product matching, businesses increase their sales by suggesting additional product options for their clients.
For instance, a buyer looks for remote control for a particular device. They put the chosen item into their shopping cart, and the system suggests buying batteries. The recommendation is relevant for the client, and they purchase both products. Instead of earning $50, a store earns $55. Multiply this figure by the number of your clients, and you will see how much your profit can increase.
Establishing Competitive Pricing
As mentioned above, the price comparison model brings advantages not only for a customer. E-commerce businesses also benefit from using deep learning in the formation of their product pricing. Machine learning enables them to automate pricing analysis and track tendencies of how prices change over time.
With price intelligence, a business can increase profit by at least 9%. Manual solutions are not that effective in this regard because they require human-based research and manual data extraction. It slows down data processing and reduces its value. Moreover, if done manually, pricing analysis impedes business scalability since the process depends on time and resources.
Improving Product Listing
With deep learning in ecommerce, you can improve the way your products are listed. Technology-powered analytics allows you to learn more about client behavior and understand how to better rank products in the list.
You can take matched stores and drive the analysis of competing listings. Trained algorithms on competing listings will provide you with insights on the way products are described and titled. Analyzing different sections of identical products helps to detect gaps in presenting an item For instance, detecting the right keywords may increase the chances that your offer will be ranked high in the overall list. As a result, you digest the given output and get the tools to increase conversion.
Detecting Copyright Infringement
Ecommerce businesses spend thousands of dollars on building a solid brand identity. They create unique visual and verbal content, write catchy product descriptions, develop attractive designs. Surely, they do this with the aim of differentiation and standing out from the competition. Keeping an offer unique is of the utmost importance for retailers. This is why prompt detection of copyright infringement is crucial in e-commerce.
We can apply the image similarity model to detect cases when a company’s design is stolen by another business. To reinforce the results, the title similarity model can also be used. Using titles as an input, we facilitate the search process and adjust the output. Combining the title similarity model with image increases the chances that businesses will promptly detect copyright strikes (even if the information presented along with design significantly differs from the original).
ML allows for precise product matching, making ecommerce development more consistent. For a seller, it means a higher profit and a clear view of their competitive power. With the help of machine learning algorithms, they prevent situations when a client is presented with an endless list of duplicated products. Sellers provide users with a chance to opt for the most advantageous offer, which pays them off in increased trust and client loyalty.
Also, ML helps to organize identical product offers in a logical and rational way. It turns e-commerce marketplaces into powerful, intuitive, and convenient spaces for doing business that satisfies the needs of the most demanding clients. If you want to integrate ML into your ecommerce platform, contact Forbytes. Our dedicated ML professionals will gladly help you to reach your business goals and make the most of machine learning power.