Product categorization in e-commerce platforms is an action of placement and organization of products into their respective classes. It has attracted a lot of research interest and attention as one of the most fundamental tasks in e-commerce. Categorization has high importance for both buyers and sellers, and plays a significant role for various downstream tasks such as product search, price comparison, and complementary product recommendation. In this work, we study product categorization based solely on the product's title and, specifically, prefixes of the title. This aims at identifying the category at an early stage of the selling flow, the process in which a seller uploads an item ofered for sale to the e-commerce platform. Once the item's category is identified, the rest of the selling process can be adapted accordingly and expedited towards a smooth conclusion. We perform an extensive analysis of title prefix categorization, inspecting to what degree the product categorization task could be efectively accomplished while only using the beginning of the title. To this end, we propose BERT-Attrs, an extension of BERT that considers, in addition to the prefix's representation, also the association of its tokens with attributes, such as brand, color, or material. Evaluation, conducted over datasets from two of the world's largest e-commerce platforms, with hundreds of categories, considers the prefix-based categorization task both from a classification and recommendation points of view. To the best of our knowledge, we are the first to introduce and study the task of product categorization based on title prefixes.
E-commerce platforms have demonstrated tremendous growth over the past decade, with the
number of products available for online shopping rapidly increasing year over year. Maintaining
these products in an organized manner requires much efort both from the sellers who upload
their inventory for sale and from the e-commerce platforms hosting these products. Sellers are
expected to provide as accurate product data as possible, which in turn has a major efect on
buyers’ purchase likelihood and the marketplace’s success as a whole [
When sellers upload a new product for sale on an e-commerce platform, they need to provide
much information, so that it can be correctly and attractively presented to potential buyers.
Often, the upload process involves coming up with a product title, selecting the most relevant
category in the platform’s product taxonomy, deciding which product attributes to provide,
uploading images, writing a product description, and providing information pertaining to price,
shipping, and available quantity in stock. We refer to this process, supported by a dedicated
user interface in leading e-commerce platforms, as the selling flow (also sometimes referred to
as the listing flow) [
CEUR Workshop
ISSN1613-0073
to the specific taxonomy and terminology of the e-commerce platform, come up with the best
title for their product, and verify that the most important attributes are provided, to ensure that
the product is easy to find, highly ranked by recommendation algorithms, attracts the buyer’s
attention, and is perceived as a quality purchase [
A first step that can play a key role in making the selling flow smoother is the identification of
the correct category of the product ofered for sale. Once the ofered product can be associated
with a category or type (e.g., a wristwatch, a juicer, or a backpack), the selling flow can be
adapted to facilitate the remainder of the process for the seller and enable a swift and productive
conclusion. Specifically, once the category is known, the seller can be directed to provide the
most relevant information for that category. For instance, the platform can explicitly ask the
seller to provide the name of the network carrier and the capacity of the internal storage if it
categorizes the product as a smartphone. In other cases, the platform can redirect the seller to
a separate flow if it detects the specific product category. For example, a platform for selling
consumer electronics may not support personal computers and therefore redirect sellers to an
afiliated website when it detects that they try to sell a laptop. Furthermore, domain-specific
tools can be applied on top of seller-provided information once the product’s category is known.
For instance, computer vision models for identifying product attributes that specialize in specific
business verticals (e.g., Jewelry, Toys) can be applied to seller-provided images [
Automatic product categorization is one of the most fundamental tasks in e-commerce and
received attention in diferent studies over the last two decades. Many works addressed this task
under varying assumptions and product facets as the sources of data. Several works use product
descriptions and reviews (e.g., [
It should be noted that manual product categorization (i.e., when sellers explicitly assign a
category to the product at the beginning of the selling flow) is quite dificult and sometimes even
unfeasible. First, the seller should be familiar with the platform’s set of categories (which may
consists of hundreds and sometimes thousands of various categories). Moreover, the category
structure and hierarchy may difer from one platform to another, making the process even more
cumbersome and confusing. It is therefore not surprising that all large e-commerce platforms
apply an automatic approach for product categorization [
Figure 1 shows an example for the potential operation of early categorization as part of the
selling flow on eBay, one of the world’s largest e-commerce platforms. This example considers
a seller who intends to upload a camera and comes up with the title Sony Black CyberShot
W830 20MP 8x Zoom Compact Digital Camera. This title contains 10 tokens, however the
product category (digital camera) is mentioned only at its end. As the seller starts to type the
title, it is dificult to identify an accurate category based on the first two tokens, since Sony is a
manufacturer of many types of electronic goods (e.g., TVs, Cameras, Smartphones) and Black
is a common color across a wide variety of categories within the Electronics business vertical.
However, the third token (CyberShot) is a well-known series of digital cameras, therefore already
disclosing the product’s type without an explicit mention. This early category recognition allows
for applying an attribute extraction model [
To our knowledge, all prior eforts on product categorization considered the product title as a
whole [
The task of early categorization based on only very few tokens is challenging, since the input
is short and partial. Our solution is based on state-of-the-art natural language processing
methods, which involve contextual embeddings of the prefix tokens produced by language
models. To allow more eficient category inference, we extend BERT [
Our work’s main contributions can be summarized as follows: • To the best of our knowledge, we are the first to introduce and study the task of product categorization based on title prefixes, motivated by the need to enhance the selling flow experience on e-commerce platforms. • We provide an extensive analysis of the efect of title prefix length on product categorization quality, also tying to the length of the original title. • We propose a novel method for product categorization that extends BERT to include product attributes, which we empirically show to consistently improve performance. • We address the product categorization task as a recommendation problem, which can be an integral part of the selling flow, and show that this approach yields high performance gains, making it applicable for short prefixes, with as few as 3 to 4 tokens.
E-commerce platforms provide a space for buyers and sellers to connect and engage in online
transactions. While much research has been conducted to improve the experience of buyers,
for instance in product search [
Our work examines the task of product categorization, motivated by the desire to enhance
the selling flow as early as possible, by adapting it to the type of the product ofered for sale.
To this end, we take a novel approach by using only the first few tokens of a title to predict
the product’s category, rather than relying on the whole title or additional information such
as description, attributes or images, as has been done in previous work (e.g., [
In general, the task of product categorization in e-commerce has received a lot of attention
in the last two decades. Several works used product descriptions to identify a product’s
category [
A few works used solely the title for product categorization, since it is typically available
for all products. Shen et al. [
A related body of research to our work focuses on short text classification. The core challenge
of this task is to address the lack of available information in short texts. A number of works
proposed methods that employ external information, such as latent topical representation and
taxonomy-based semantic features, to enhance the representation of the short texts [
and methods we use to address it.
In this section, we describe the product categorization task we examine, as well as the models The task of product categorization aims to identify the category of the product given its characteristics. In our work, the source of information taken into account for conducting the task is only the product’s title, and specifically a prefix of the title. We examine categorization in its classic approach, as a classification problem. To further improve performance in a way that would be applicable early in the selling flow, we also suggest and evaluate a recommendation approach, which entails the suggestion of the top-k most probable categories. Both approaches require a learned model and the recommendation approach also requires the cooperation of the seller in selecting the correct category out of a short list of suggested categories during the selling lfow. The benefit of the recommendation approach is that it allows more room for error, since identifying the correct category among the top-k is suficient. On the other hand, classification can facilitate an automatic flow without any seller involvement (e.g., via an API that allows batch listing of multiple items). It is therefore interesting to examine the trade-of between these approaches.
loss:
Our experimental setting focuses on analyzing the performance of models based on the
pretrained BERT architecture [
In addition to diferent methods utilizing BERT, we also inspect the results of LSTM, a recurrent neural network based on long short-term memory architecture [52], with pre-trained word2vec embeddings using continuous bag of words (CBOW) with negative sampling [53].
Throughout this work, we considered prefixes of length in the range of ∈[1, ⌊ title length in tokens. For instance, for titles of length 12, we examined prefixes of length 1 to 6 tokens. We inspected two diferent learning methods for prefix categorization. The first trains only on complete product titles and the second trains on prefixes. Specifically for the latter, for each title in the training set, in each iteration of the training process (epoch), a prefix length is 2 ⌋], where is the drawn uniformly at random out of the [1, ⌊ ⌋] range and the corresponding prefix is used for training.1
As we will later show, prefix-based training substantially outperformed training based on complete titles. We therefore proceeded to build our models based on title prefix training. Since title prefixes consists of very few tokens, we sought to extend their representation with additional information to allow more efective learning, as described in the next section.
In addition to the standard BERT-based approach (training using randomized prefixes), we devised an extension of BERT, termed BERT-Attrs, which makes use of novel input enhancement methods. As will be shown in Section 5, BERT consistently outperformed LSTM when trained on title prefixes, and we therefore opted to focus on a BERT-based extension. Concretely, BERTAttrs leverages additional information regarding the attributes of a product, extracted using a state-of-the-art named entity recognition (NER) method specialized and trained for attribute extraction from product titles [54]. This NER method aims at associating each token in the title prefix with an attribute name such as brand, color, or size2. Overall, 63.68% of the tokens in the training set could be associated with an attribute name. The extracted attribute information is then injected into the BERT input as an additional sentence in the form: “[ 1] value1 [ 2] value2 [ ] valueN”, where each [ ] is defined as a special token. For example, for the prefix:
“green adidas cotton” the derived “attribute tokens” sentence would be:
“[Color] green [Brand] adidas [Material] cotton”.
In cases where prefix tokens could not be associated with an attribute name, BERT-Attrs injects them along with the special corresponding token [UNKNOWN] as their attribute name. Considering the example above, suppose the last token in “green adidas cotton” could not be classified into an attribute name, BERT-Attrs would inject the sentence:
“[Color] green [Brand] adidas [UNKNOWN] cotton”
Based on this injection, the whole input to BERT is of the form:
[]
<prefix tokens> [ ]
<attribute tokens> [ ]
where [CLS] and [SEP] are standard tokens, used as in the original BERT [
In addition to the methods mentioned above, we examined several other representations that consider attribute information as part of the input to BERT. First, we considered a variant of BERT-Attrs that excludes tokens that cannot be associated with an attribute from the attribute injection altogether. In other words, only tokens that can be associated with an attribute name are included in the attribute-based representation that follows the original prefix. Second, we experimented with a variation that aims to exploit BERT’s pre-trained language understanding by adding the attribute information as natural language to the prefix. For instance, the attribute 1Wtiteleaolsfoleenxgptehri miennteeadchwtitrhainainvagriitaetrioantioonf.thTishiaspvparroiaactihonth,ahtowtreaviners, opveerrfoarlmlperdefixsiemsiilnartlhyearnadngoeften[1,s⌊li 2g⌋h]tfloyrmeaocrhe poorly compared to drawing a prefix at random, while bearing substantially higher computational costs. 2We used an in-house NER model trained on eBay’s product titles. The quality of the model is depicted at [54]. injection of our example prefix would be in the form: “The prefix green adidas cotton contains green as [Color], adidas as [Brand] and cotton as [Material]”. Finally, we considered an approach that operates directly on BERT’s architecture. This is done by adding the word embedding of the corresponding attribute to each token’s representation. Each token’s embedding is therefore derived by:
⊕ ⊕ ⊕
where ⊕ marks the tensor addition operation; WordEmb, TokenTypeIDEmb, and PositionEmb
are the standard components in the BERT architecture [
All methods described in the previous paragraph consistently underperformed both BERTAttrs and BERT-AttrsOnly. We therefore exclude them from our reported results, for clarity of presentation, and focus on comparing LSTM, BERT, BERT-Attrs, and BERT-AttrsOnly, trained over title prefixes as described in Section 3.3.
In this section, we provide details about our setup for experimentation, including the two datasets and their basic characteristics and evaluation metrics. We publicly release our implementation for reproducibility purposes.3
We experiment with two diferent datasets. Our main dataset is from eBay, one of the world’s largest e-commerce platforms.4 For reproducibility purposes, we also include results over a 3https://github.com/titleprefixes/prefixes_code 4A small sample of eBay’s data can be found at https://github.com/titleprefixes/prefixes_code/blob/main/data/ebay_ example_data.tsv public dataset – the Amazon products dataset. In both datasets, the categories are defined by a taxonomy , where verticals are at the highest level and the categories we consider for the task are at the lowest level.
This dataset includes a sample of 17 million listing titles and their corresponding categories from eBay’s logs. The sample contains listed items (listings) that were ofered for sale on the United States site during December 2020. In this dataset, the listings stem from 6 diferent verticals and 704 corresponding categories. We randomly split the data into training, validation, and test sets with a ratio of 60%, 20%, and 20%, respectively. Table 1 presents several examples of titles and their corresponding verticals and categories in the dataset. In some cases, the diferences between categories are subtle. For example, Watches Parts and Watches Accessories or Golf Equipment and Golf Parts Repair. These minor diferences and overall large number of categories entail dificulty in the categorization task.
We consider the Clothing, Shoes and Jewelry vertical (also referred to as “Fashion”) of the publicly available product dataset from Amazon [55], another one of the world’s largest e-commerce platforms. Overall, this portion of the dataset contains 464,745 product titles spanning 147 categories. As in the eBay dataset, some categories are rather similar to one another, for instance Men’s Shirts and Women’s Shirts or Men’s Wrist Watches and Men’s Pocket Watches. We use the same splitting scheme of 60%-20%-20% for training, validation, and test sets, respectively, as for the eBay’s dataset. The products in this dataset span the years 1996-2014. Table 1 presents several examples of titles and their corresponding categories in the dataset. Basic characteristics of both the eBay and Amazon datasets are summarized in Table 2.
the number of titles in the dataset associated with that category). These statistics convey the extreme imbalance with respect to class size in both datasets. In the eBay dataset, 10.79% of the categories (up to 100 titles per category) cover only 0.02% of all titles, while in the Amazon dataset, 29.25% of the categories (also up to 100 titles per category) cover 0.2% of all titles. On the other hand, on eBay, 4.54% of the categories (more than 100 titles per category) account for 62.6% of all titles, while in Amazon 6.12% of the categories (bin “10K-100K”) cover 42.71% of the titles. Naturally, this imbalance stems from the skewed distribution of e-commerce items across categories, reflecting high diferences in their popularity. eBay dataset, the average length is 12.24, with a standard deviation of 3.59 and a median of 12. In the Amazon dataset, the average length is 11.06 tokens, with a standard deviation of 4.24, and a median of 11. The titles on the eBay’s platform are strictly restricted to 80 characters, whereas on Amazon there is a restriction of 200 characters. Thus, in rare cases, Amazon’s titles can become substantially longer when compared to eBay. This phenomenon at the tail of the distribution can be observed in Figure 2.
For the category classification task, we report accuracy, macro precision, and macro recall [ 56]. We focus on these metrics to understand the model’s performance in the scenario of extreme imbalance. The macro metrics consider the average precision and recall across all classes, while assigning each class with an equal weight, regardless of the number of its associated instances. We omit micro precision and recall as they are both equal to accuracy when each data point is assigned to exactly one class. Formally, the metrics are defined as: =
= ∑=1 1 ̂ = .
1 =1 ∑ where is number of instances, ̂ is the predicted label of instance , is its true label, is the number of classes and ∈{ , }
For the category recommendation task, we focus on the Hits@k metric [57, 58], defined as the fraction of examples where the correct class is among the top-k ranks. It is worth noting that Hits@1 is equivalent to the accuracy metric. Formally, it is defined as:
∑=1 1 ∈ ({ ̂ }=1 ) where
is the set of highest-ranked classes according to model’s predictions.
In this section, we report results for classification and recommendation, using the methods
described in Section 3. We start by showing that the results produced by training using complete
titles are not satisfactory, and thus motivate our approach of training over prefixes. We then
compare diferent approaches for the prefix-based product categorization task using prefixes for
training, for the problem of classification, followed by analogous results for recommendation.
Finally, we examine the contribution of diferent attributes to the categorization performance,
using ablation tests. The majority of the results are reported over a test set of titles with an
original length of 12 tokens, which is the most common across our two datasets combined, as
described in Section 4.1. We report results for all prefix lengths in the range of [
We first examine the use of complete titles as training examples for prefix-based categorization. To this end, we apply the BERT and LSTM models trained over complete titles on our varyinglength prefixes. The classification results for titles of length 12 are summarized in Table 4. Overall, we observe low values with respect to all reported metrics and datasets. Even for prefixes of length 6 (half of the title), the accuracy did not exceed 74% and 60% over the eBay and Amazon datasets, respectively. However, when applied on complete titles, the accuracy of both models exceeds 90% on both datasets. This indicates that the learned patterns over the complete titles do not generalize well when used to categorize prefixes. The setup of our proposed methods is intended to address this issue and our observations provide empirical support.
In this section, we evaluate our ability to classify title prefixes using our proposed methods, discussed in Section 3, and explore the trade-ofs between prefix length and categorization performance. Tables 5 and 6 present the classification performance results when training on prefixes using the LSTM, BERT, BERT-Attrs and BERT-AttrsOnly methods over the eBay and Amazon datasets, respectively.
As expected, classification performance degrades as prefix length decreases. For prefixes up to 3 tokens, we observe poor classification performance across both the eBay and Amazon datasets. This reflects the dificulty in identifying the correct category based on very few tokens. Consider, for example, the titles presented in Table 1 in the Fashion vertical. In both datasets, we observe titles that share the same prefix when considering the first 3 tokens (i.e., Calvin Klein Womens or The North Face), but are from diferent categories. Therefore, even a perfect classifier would not distinguish between them based on three-token prefixes. On the other hand, for prefixes of length 6, i.e., 50% of the title’s tokens, we can observe much improved performance. In particular, over the eBay dataset, all BERT-based methods (BERT, BERT-Attrs and BERT-AttrsOnly) reach accuracy of over 80%.
Examining Table 5, it can be observed that the BERT-Attrs method outperforms all other methods in terms of accuracy across all diferent prefix lengths. It also outperforms the other methods in the vast majority of the cases in terms of macro precision and macro recall. The consistent gap in performance between BERT-Attrs and BERT indicates that modeling the attribute information as part of the input prefixes helps the categorization process. The uplift of BERT-Attrs on top of BERT is more substantial for short prefixes (e.g. +5.7% for prefix length 1 compared to 0.8% for prefix length 6), which are the most dificult to categorize and therefore can benefit the most from the additional information used by the BERT-Attrs model. The BERT-AttrsOnly model yields consistently lower performance than the BERT-Attrs model (in accuracy), indicating that the inclusion of the title in its original form as part of the input is not redundant. The results over the Amazon dataset in Table 6 show similar trends, with BERT-Attrs outperforming the other methods across all prefix lengths except for length 1 where BERT-AttrsOnly performs best. The gap between BERT-Attrs and BERT is consistent, while largest for prefix lengths of 3 and 4, and substantially diminishing for the longer prefixes of 5 and 6 tokens.
Figure 3 presents the classification accuracy as a function of the original titlle length (for
original length ∈[
Prefix len = 4
Prefix len = 5
Prefix len = 6 100 95 y90 c rau85 c cA80 75 70
In this section, we examine the performance of the learned models in the scenario of category recommendation for title prefixes. That is, we use our models to recommend the top-k most probable categories based on the input. This is achieved by considering the scores of our classifier as ranking scores over categories and using the induced ranking for the recommendation.
Table 7 presents the Hits@k results of all four methods over both the eBay and Amazon datasets for titles of length 12. We focus on ∈{3, 5} , as we assume these values allow sellers to traverse the candidate list and select the suitable category without substantial cognitive load. We also focus on low values of to allow the seller to view the entire list of categories on a single screen.
Observing the table, we can see that as in the classification results, performance improves with the length of the prefix. Moreover, as in the classification case, the BERT-Attrs model outperforms all other models across all prefix lengths for the eBay dataset, and in the vast majority of the cases for the Amazon dataset (in a few cases BERT-AttrsOnly slightly outperforms BERT-Attrs). The gap between BERT-Attrs and BERT is consistent for both Hits@3 and Hits@5 across all prefix lengths over both datasets, and is particularly large for the shorter prefixes, where attribute information is most essential to allow narrowing down the list of potentially matching categories. Overall, these results reinforce the added value of modeling attribute information. The small but rather consistent gap between BERT-Attrs and BERT-AttrsOnly attests again to the benefit of including the original title as part of the input representation. Also consistent with previous findings is the lower performance of the LSTM model compared to the BERT-based models.
In contrast to the classification case, however, the performance using all four models reaches high values starting already from very short prefixes. For example, considering the best performing BERT-Attrs model, while its accuracy over the eBay dataset only exceeds 82% for a prefix length of 6 (Table 5), it exceeds 82% with Hits@3 when the prefix length is only 3 tokens (25% of the title) and exceeds 81% with Hits@5 when the prefix length is only 2 tokens. This indicates that while short prefixes tend to be ambiguous and may fit multiple categories, the cardinality of the overall set of potential matching categories is already low after 2-3 tokens, in many cases. This enables the recommendation approach, which loops in sellers and allows them to disambiguate at an early stage, to be highly productive. In Table 1, the provided examples in the Electronics vertical demonstrate the above point. For example, Sony Playstation 5 (or PS5) is a prefix that can be shared across diferent categories, however, the list of relevant categories is limited to the video games domain, with only few alternatives.
Table 7 indicates that by the time a seller has typed 5 of the title’s tokens, the Hits@k values using BERT-Attrs over both datasets are above 90%. For prefix length of 6 tokens, Hits@3 on both datasets is higher than 92% and Hits@5 exceeds 95.5%, reaching as high as 96.14% over the eBay dataset.
To conclude this section, we explore in more depth the potential gains in performance applying
the recommendation method with diferent values of . We focus on short prefixes of length 2, 3,
and 4, as these demonstrated rather low accuracy using the classification approach (Tables 5
and 6). Figure 4 plots the Hits@k as a function of ∈[
Our BERT-Attrs model relies on attribute information to learn diferent patterns within the titles to perform categorization. We therefore set out to explore the influence of key attributes on the categorization task using ablation tests. In this experiment, we focus on three diferent verticals within the eBay dataset – Fashion, Electronics, and Home & Garden. For each vertical, we consider the most popular category within the eBay’s test set (see Section 4.1), as measured by the number of associated listings. The categories explored are Shirts, Video games, and Home Decor from the Fashion, Electronics, and Home & Garden verticals, respectively. For each category, we consider all its associated titles in our test set of at least 12 tokens and their corresponding 6-token prefixes. The total number of such prefixes in our test set from each category is 236,223, 39,757 and 47,623, respectively. For each category, we use the NER model [54] to extract attribute values from the prefixes. We report results over the top 7 most frequent attributes in each category. For each such attribute, we consider all prefixes in the corresponding category that include it. We measure the categorization accuracy across these prefixes, and compare it to the accuracy over the same set of prefixes while removing all tokens that correspond to values extracted for that attribute. This allows us to measure the relative impact of removing diferent attribute values from the prefix on the categorization performance.
Table 8 reports the accuracy diference between prefixes that include and exclude each of the top 7 attributes in each of the categories. Alongside the accuracy diference, which attests to its importance to categorization performance, the average length (in tokens) of each such attribute across all category prefixes that include it, is presented. Observing the table, it can be seen that the type attribute is the most important for Shirts and Home Decor categories. In the Shirts category, the average length of the type values is only 1.38 tokens, ranking only 5th out of the seven attributes in terms of length, indicating that the contribution to performance does not necessarily coincide with length. In the Home Decor category, the average length of type is 1.93 tokens, which is the third highest. The importance of type to the categorization task is intuitive, as it is typically closely associated with the category. For example, possible values for the attribute type in Shirts and Home Decor categories are T-shirt and Wall Picture, prefix len = 2 prefix len = 3 prefix len = 4 1 2 3 4 5 6 7 8 9
10 respectively. These values are associated strongly with their respective categories. We note that type attribute tends to appear towards the end of the title. For instance, in the Fashion vertical of the eBay dataset, it appears on the second half of the title in 65.5% of its occurrences.
Other than type, the style attribute (e.g., retro, loungewear) is found to be the most important attribute for categorization in the Shirts category. In Home Decor, the model attribute is as nearly important for categorization as type, with values such as Cierra by Uttermost and ST1216B by FAIRFIELD.
For the Video Games category, the most important attribute is game name. This is an example of an attribute that is mostly unique to a specific category (e.g., Tony Hawk’s Pro Skater) and therefore its occurrence is highly revealing. It also tends to be exceptionally long at over 4 tokens. Therefore, its removal leaves very few tokens on the prefix, leading to the very large performance gap in its ablation test. The type attribute, which is most important in the other two categories, tends to be more generic in Video Games: its most common values are Disc or Game. Accordingly, its importance as reflected in the ablation test is lower.
The commonly-used brand attribute is not found to be among the most important attributes for the categorization task, ranked fourth for both the Shirts and Home Decor categories. The relatively low importance of the brand attribute in Shirts is intuitive, as the same brand can span diferent categories. An example of such phenomenon is the known brand of Michael Kors, which has products in various categories, including dresses, shoes, shirts, handbags, and even watches. However, for the Home Decor category, the brand attribute can be more distinctive. For example, it can be associated with the designer’s name (or an artist’s name), which is less likely to be shared across diferent categories. The brand attribute is particularly common at the beginning of the title: in the eBay dataset, it appears on the first half of the title in 88.6% and 83.2% of its occurrences in Fashion and Home & Garden, respectively.
We studied the task of product categorization based on title prefixes. To our knowledge, we are the first to introduce and explore prefix-based categorization, which can play an important role in e-commerce platforms’ selling flow, allowing to adapt early to the unique characteristics and attributes of a specific category, out of hundreds of available options. We first demonstrated that using both BERT and LSTM models, trained on complete titles, for the prefix categorization task, results in poor performance, and hence suggested a simple approach for training a model based on title prefixes, where for each title a prefix of random length is selected in each iteration of the training process. This method accomplished substantial performance improvements, with an accuracy of roughly 80% for prefixes of half the length of the title (e.g., 6 out of 12 tokens), and BERT consistently outperforming LSTM. Following, we suggested BERT-Attrs, a novel yet simple extension of BERT for learning over title prefixes, by extending the representation of the prefix with attribute information. This extended representation showed to yield consistent performance enhancement, especially over prefixes of short length, where the original information provided in the prefix is more limited. Interestingly, our results indicated that classification performance using our prefix-trained approach is highly dependent on the prefix length, naturally increasing as the prefix is longer, but is almost completely oblivious to the length of the original title, for a given prefix length.
We demonstrated the dificulty of prefixed-based classification, which often introduces ambiguity, as the first few tokens are not uniquely identifying the specific category. Our attribute ablation tests showed that many attributes that commonly occur on titles, such as color, size, material, and even brand, which often appear at the beginning of the title, do not contribute much to accurate categorization. On the other hand, a revealing attribute such as type commonly appears toward the end of the title. This ambiguity, however, typically narrows down to only a handful of categories after as few as 2 or 3 tokens have been provided. We therefore suggest addressing the task using a recommendation approach, which loops sellers in the process and allows them to select the category out of a shortlist of candidates. We showed that this approach can dramatically improve the performance, reaching a hit rate of over 95% over both datasets for prefixes of 6 out of 12 tokens, when shortlisting to 5 candidates. Moreover, for prefixes of 3 tokens only, the hit rate within the top 5 already exceeds 88%. The category selection out of the recommended shortlist is, thus, a key step in our envisioned selling flow. It enables quick identification of the type of product ofered for sale, and adaptation of the remainder of the process towards a simplified and smooth conclusion.
Our results quantitatively draw the trade-ofs between both the prefix length and the size of the candidate shortlist, and categorization performance. These suggest a variety of alternatives for e-commerce platforms to implement prefix-based categorization, considering how early they want to provide a suggestion, whether they want to allow sellers to select the category out of a certain-sized list, and how accurate they desire the results to be. All of our results generalize similarly across two of the world’s largest e-commerce platforms, eBay and Amazon, and can be reproduced over the public Amazon dataset.
As the task of prefix-based categorization has not been previously studied, it ofers many opportunities for future research. Validating the results on additional e-commerce platforms, with diferent types of category sets and granularity, can help to further generalize. Even more importantly, in-vivo experimentation with our proposed solution is necessary to quantify its impact on the simplification and improvement of the selling flow. This experimentation would also provide an opportunity to leverage otherwise unattainable user feedback, such as whether and which suggested category was selected, at what stage, and how the selection was reflected in the remainder of the selling flow. Using prefix-based categorization in an online setting may also influence the way titles are formulated. Sellers may become increasingly aware of this feature and adjust their titles to contain category-distinctive tokens, such as type and model, in the beginning, to facilitate faster categorization. Finally, experimentation with additional models, including the potential use of LLMs within the flow, is an interesting future direction. Both in terms of the potential to increase performance and in terms of the challenge of applying large models in an online scenario.
The authors have not employed any Generative AI tools. books, in: Proc. of ICCV, 2015, pp. 19–27. [52] S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural computation 9 (1997) 1735–1780. [53] T. Mikolov, K. Chen, G. Corrado, J. Dean, Eficient estimation of word representations in vector space, arXiv:1301.37810 (2013). [54] Y. Xin, E. Hart, V. Mahajan, J.-D. Ruvini, Learning better internal structure of words for sequence labeling, arXiv:1810.12443 (2018). [55] R. He, J. McAuley, Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering, in: Proc. of WWW, 2016, pp. 507–517. [56] M. Grandini, E. Bagli, G. Visani, Metrics for multi-class classification: an overview, arXiv:2008.05756 (2020). [57] L. Yao, C. Mao, Y. Luo, Kg-bert: Bert for knowledge graph completion, arXiv:1909.03193 (2019). [58] Y. Tay, V. Q. Tran, M. Dehghani, J. Ni, D. Bahri, H. Mehta, Z. Qin, K. Hui, Z. Zhao, J. Gupta, et al., Transformer memory as a diferentiable search index, arXiv:2202.06991 (2022).