At the time of our intervention, Wallapop’s categorization experience faced a multitude of long-standing challenges. These challenges had remained unaddressed for a considerable period, adversely affecting the accuracy of categorization. Consequently, the discoverability of items for prospective buyers was compromised, leading to obstacles in achieving successful transactions.
Of particular note, when users conducted searches within specific categories and subcategories, we observed a decline in success rates when compared to unfiltered searches. This observation strongly hinted at an underlying data accuracy problem that was evidently impeding the overall performance of the platform.
In our data analysis, we observed that a significant portion of searches, comprising 68% of the total, were conducted without specifying a category. Interestingly, searches within specified categories had a slightly lower performance, with a 5% decrease in Search to Purchase Intention (PI).
Furthermore, only a small fraction, 4.4%, of searches included a subcategory filter. Notably, searches with filters applied demonstrated diminished conversion rates, with a decrease of 12.8% in Search to Click and 13.1% in Search to PI when compared to searches conducted without any filters. This trend of lower conversion rates was consistent across nearly all categories, as indicated by our analysis.
Beyond our focus on buyer success, we also recognized at that time that sellers faced significant challenges when uploading items. They encountered difficulties in selecting an appropriate category or subcategory, often feeling that the available categories were not relatable or accurate. Many sellers found the existing category structure too abstract, and they believed that it did not align with users’ mental models.
How might we facilitate and direct sellers in categorizing their items appropriately to boost their success on our platform?
Furthermore, sellers voiced complaints about Wallapop changing their chosen categories without providing adequate information or explanations.
The Flow
These combined issues highlighted the critical need for improvements in Wallapop’s categorization experience.
Illuminating the root causes behind this lack of accuracy, we have identified several elements within our product that have contributed to this issue.
One crucial factor is the presence of a model that was making corrections without offering any feedback to the user. This lack of transparency hindered users from understanding and learning from the corrections made, exacerbating the problem.
Another challenge stems from a fragmented user experience, where users were required to select categories and subcategories through separate flows. This disjointed approach not only created confusion but also led to inconsistencies in the categorization process and hindered users’ ability to seamlessly navigate the entire category tree.
Furthermore, our model’s limited consideration of subcategories, focusing solely on overarching categories, has compounded the problem. Neglecting the finer granularity of subcategories, which is often crucial for accurate categorization, has introduced inherent limitations in our categorization system.
The Approach
Transitioning to a Suggestive Model
The existing system, which corrected users when they categorized items incorrectly, presented several issues:
- Lack of User Awareness: Users often remained unaware that their categorization errors were being corrected, leading to confusion and frustration.
- Limited Engagement: The system’s corrective approach discouraged user engagement and participation.
The first step involved transitioning from a corrector to a suggester. By doing so, we aimed to empower users to make decisions with greater autonomy while eliminating friction in the user experience.
To accomplish this goal, we embarked on a journey to comprehend the key input factors that guided the model’s decision-making process. Our objective was to ensure that we not only supplied these inputs effectively but also prepared the model to provide accurate suggestions. In this particular scenario, the critical input factor was the item title. To test our hypothesis, we initiated an experiment by repositioning the title as the initial step in the workflow, followed by the category suggestion process.
This experiment has substantiated our hypothesis: suggesting a category instead of simply correcting it is instrumental in assisting sellers to accurately categorize their items. This enhancement not only improves the overall cohesion of the uploaded items but also enhances the conversion rate within the workflow. Indeed, we witnessed a tangible increase in conversion rates, signifying our success in simplifying the process for sellers.
Nevertheless, it’s noteworthy that this improvement hasn’t yielded a discernible impact on the subsequent success of these items with buyers. While the categorization process has become more efficient, the correlation with buyer engagement and transactions remains an aspect requiring further investigation.
Extending Suggestion System to Subcategories
Our pursuit of improvement led us to address the aspect that remained untouched by the previous experiment, specifically the Upload to Purchase Intention (PI) phase. To enhance this aspect, we made the strategic decision to expand our model’s capabilities to include subcategory suggestions. We believed that by suggesting subcategories, we could further assist sellers in accurately categorizing their items, thereby enhancing the coherence of the uploaded inventory.
In this initiative, our primary focus was on providing subcategory suggestions while retaining the existing fragmented flows. We seamlessly integrated these suggestions into the existing workflow.
Achieving this required us to retrain the model using a new dataset comprising subcategories from across our entire catalog. This pivotal task was undertaken by our dedicated Data Scientist, who ensured that the model was adequately prepared for this crucial enhancement.
As a result of this experiment, we not only continued to enhance the flow’s conversion rate but also achieved a notable 6% increase in the success of these items on our platform. By making them more discoverable to potential buyers, we have effectively improved their visibility and desirability.
Combining Category and Subcategory Selection Into One Flow
To effectively address the remaining challenge, we’ve made the strategic choice to consolidate and harmonize the selection processes for both categories and subcategories.
By merging these two flows into a cohesive flow, we aim to simplify the user experience, improve efficiency, and create a more seamless solution. This integration not only optimizes the user journey but also enhances our ability to manage and analyze data, enabling us to make more informed decisions and ultimately deliver a more robust and user-friendly product.
The Conclusion
In conclusion, our journey to improve the accuracy of item categorization and discoverability on our platform has been marked by significant strides. We initiated experiments that not only simplified the categorization process for sellers but also led to notable increases in conversion rates. By suggesting categories and subcategories, we have empowered sellers to place their items more accurately, improving connectivity within our platform. This, in turn, has not only increased the success of these items but also enhanced the overall experience for both sellers and buyers.
While there is always room for further refinement, these initiatives represent a significant step forward in our ongoing commitment to delivering a seamless and rewarding user experience.