Final

Abstract

Despite the popularization of online clothes shopping and social networks, existing internet communities do not fully leverage the advantages of both systems. Suit Up is a new social network that attempts to create a practical and enjoyable social space centered around clothing items. With an account, users can explore, archive, shabre and receive recommendations on what to wear. Suit Up’s bookmarking tool allows people to keep track of personal clothing items found on any website. The system recommends new items and outfits for the user based on his/her actions, including following other users, ratings and bookmarks. Other tools such as Hot-or-Not and Outfit Creator not only provide entertainment and interest in the system, but also contribute to improving the user’s item recommendations. Suit Up provides many important social objects, and aims to be appealing to people with varying degrees of interest in their clothing.

Introduction

With constantly changing weather, cultural norms and fashion trends, people are challenged with the decisions on what to wear each and every day. Through brainstorming sessions, a survey and interviews with potential users, we designed and developed Suit Up. It is a new social network that provides a way for users to learn about new clothing items and receive recommendations based on preferences. After our initial research, we conceptualized system features through sketches and whiteboard sessions. Through the Mockingbird prototyping tool, we created lo-fi mockups of Suit Up pages, which we tested with several users. We began basic functionality development, and simultaneously processed our research feedback into mid-fi prototype designs. The finalized version of Suit Up was then developed and completed with much of our originally designed functionality.

Use Cases & Motivation

Through some early ideation discussions, our team became interested in the idea of fashion and clothing social spaces on the web. Clothing and accessories have shown to be a successful concept on the web through e-commerce and online shopping, which have been steadily rising in recent years [1].

Even though clothing is prominent on the internet, it still has not found a large amount of success with social networks [2]. Part of this has to do with the ability of a system to identify contextual information for a user and the ability to turn that into value for that user (e.g. recommendations through the system). Also, fashion is constantly changing and the system must keep up with the user’s latest behaviors. We saw this as an interesting opportunity to explore how clothing and fashion can work in a social web environment.

In our brainstorming sessions, we discussed anecdotal evidence about how people struggle to figure out what to buy and wear each day, and for special occasions. We then created a survey to observe 71 peoples’ feelings toward clothing and fashion. We also interviewed several subjects in their 20’s, who we saw as potential users of our system.

According to our survey, even though 65 tend to shop online at least occasionally, there is no standard way that they are able to track potential purchases. A number of respondents also said they seek advice from sources that they trust, such as friends and family (31), but also items such as magazines (13). One user we interviewed said, “If I think someone dresses nicely then I would listen more to their advice. If I had a friend who was struggling with what to wear, their advice was less trustful. Or if they had a completely different style from me I’d trust it less.”

One inspirational example we explored from brainstorming is: “business casual” may mean entirely different things depending on the region, industry, company culture or weather. Our hope was that a new system would be a useful tool that also provided meaningful advice and recommendations in the context of a particular user’s tastes. We prepared a user scenario to also illustrate our concept.

Competition

There are a number of emerging social systems pertaining to the subject of fashion. Some sites cater to haute couture, where its users have a strong interest in unique styles. These include sites like chictopia.com and lookbook.nu, which are communities that rate looks based on users sharing their uploaded content. Other systems, such as Fashism, focus on a simple game concept where people compare two items against each other. Svpply is a social site, but is invite-only and the recommendations are not tailored to a specific user.

These communities seem to offer unique features, but generally cater to a younger audience that has a very strong passion for unique fashion trends. To some degree, this leaves people who have a more casual interest, or are simply desperate for advice, out of clothing-related social networks. The systems we surveyed also tend to focus more on entertainment, and require the user to exert the effort to seek new recommendations.

Some fashion sites provide a more traditional social networking experience, such as ThisNext. However, recommendations are left up to the users to share with each other, and very little social computation seems to contribute to the items provided to a user.

Finally, we were inspired by sites like Delicious for our clothing item bookmarking concept. We also used Shopstyle as a helpful model; it is primarily an e-commerce aggregator. It does have a social component with very active contributors. Yet, Shopstyle does not seem to have an explicit goal in terms of its social objects, which are supposed to create the bond and connection between members of networks [3]. Shopstyle allows users to do things like create outfits, join groups and friend others, but its users seem to be doing it more for posterity and with no goal for system recommendations or explicit ways to offer users fashion advice.

Our competitive analysis provided us with many ideas and a lot of context for our project, and the analysis helped us fully realize each system’s strengths and weaknesses. At this time, it does not seem that these communities incorporate mechanisms to incentivize more contribution, system recommendations, and personalization based on the user’s preferences. This leaves the user with the responsibility of creating value out of the systems. The goal with Suit Up is to provide a useful and valuable experience for people not necessarily involved in niche fashion scenes.

Implementation & Design

Suit Up was designed to mimic a user’s real life closet in addition to outfit and item selection techniques.  Our layout combined the familiar Facebook interface with the fashion necessities of closets and outfits.  We used an Apache server interpreting PHP as the foundation of the site, with a MySQL database and strong HTML and CSS templates.  The database had segregated demographic data so that it could be used later for business development purposes, linking heights or ages to certain tags or groupings of clothing items.

Figure 1. System Architecture

Figure 2. Interface Architecture

Figure 3. Lo-Fi Interface Example


Our MySQL database was appropriately normalized, although MySQL limits foreign key functionality, it did not hinder our performance.  Here is our inital database diagram.  Though its implementation has evolved slightly over the course of the project life cycle, this representation is still accurate in the basics.

One concern we have with our current architecture is the bottleneck that will surely result from saving so many images. Because links change and break as time goes on, we felt that it would be prudent to save item image files on our server, rather than hot-linking. Unfortunately, with our resources, this was our only available option. As time goes on we would procure additional storage to distribute our file load among multiple servers. Depending on the growth rate of the database we might also consider distributing the database among multiple servers. Our recommendation system, Directed Edge, will be discussed in the next section.

Computation

One of Suit Up’s primary functions is generating clothing and neighbor recommendations for users.  After considering a number of different collaborative filtering systems, we decided to use DirectedEdge because its relatively extensive documentation and PHP bindings.  Since Suit Up is currently non-commericial, we qualified for the free Developer account.

DirectedEdge does not distinguish between users and products.  Instead, it treats everything as an item.  Each item is a node that can have links which point at other items. Users, products, outfits and demographic information are the types of items that Suit Up sends to DirectedEdge.  When a user bookmarks or rates a product,  a link is created from the user to that item.  Expressing preference for an product in the Hot Or Not section of the site also creates a user-to-item link.  Similarly, when a user follows another user, a link is created from the first user to the second.  Users also link to an age range, industry, body type and height range, based on the information they provide when signing up for an account.  This enables us to immediately generate recommendations for new users, without waiting for them to rate or bookmark items.

We have implemented two types of recommendations so far: clothing, and similar users.  Recommendations are based on how strongly an item is linked to other items in its “neighborhood” in the network.


The figure above shows a simplified example of such a neighborhood. Jessie is a new user who is not yet connected to any products or users in the system.  However, she has entered information about her age and height, linking her to these demographic items. She is indirectly connected via this information to other users, such as Yvette and Brittney, who are in turn connected to products. We can guess that Jessie might be relatively more interested in the Green top than the Blue jeans or Black cardigan, and that she is more similar to Yvette than Cornelia.

Sustainability

In terms of sustainability, there are several ways that will help maintain our system:

1.    Design our own recommendation algorithm to more accurately address item suggestions.

2.    Host our own Directed Edge software to support maximum user capacity.

3.    Distribute images to another server to lighten load dependency.

Currently we stored images on our own server, but as Suit Up expands, a more distributed architecture is needed. Similarly, advertiser submitted ads would have to be stored  elsewhere. FTP accounts for verified advertisers is one option, and would allow Suit Up administrators more flexibility and oversight to ad submissions.

4.    Make a profit using the infomediary model. [4]

Companies that provide information to assist customers and/or sellers understand a given market. We would be able to serve as infomediaries those who would find this information valuable. The following are some examples that Suit Up can utilize the infomediary model:

a.    Revenue from linking to e-commerce sites.

Suit Up will be able to provide recommendations and channels to items on sites such as Amazon and ShopStyle. It would then benefit as an “affiliate” for providing links (e.g., ShopStyle pays Suit Up based on the times users click on items hosted by them).

b.    Put customized banner advertising and recommendation-like advertising on our site.

c.    Sell aggregated demographic data

Based on users’ demographic information and their rating data, retailers and marketers can analyze marketing effectiveness and thus target appropriate ads.

We also discussed other ideas for the future once we have a substantial population and revenue stream:

  • Offer a subscription for premium content (freemium model).

Initially, we provide “free accounts” to draw in as many people as possible to opt-in Suit Up. Once we have a stable user base, we might aim for a paid subscription for premium content.

Evaluation

In order to understand how our design of a clothing recommender system would make sense to people, we conducted usability testing on three potential users with our paper prototypes. The participants were specifically asked about the tasks completed on our prototype. The following points are some insightful findings:

  1. Confusing roles: all test users found ‘friends’, ‘fans’, ‘fashion neighbors’, ‘follows’ difficult to differentiate.
  2. Hot or Not: all test users liked this idea, and that guessing that was a way to introduce contribution.
  3. Confusing terms: all test users found ‘comment’, ‘bookmark’, ‘fashion neighbor’ confusing. Consistency would be critical to our labeling.
  4. Recommendations: two test users specifically expressed interest in customizing how they want to be recommended items.
  5. Search: one test user felt it made more sense to sort results by categories, or to search in “title” or “tag”. This is one feature we planned to incorporate in later development.
  6. Upload outfit: one test user admitted that she would love to see others’ outfits, but is lazy to upload hers. We needed to make sure we made the process simple and fast.

Based on the feedback we collected from paper prototypes, we created a new interface that better fits our target users’ behavior. Our plan is to first perform a heuristic evaluation to eliminate obvious and critical usability issues. Then, we will conduct usability testing on the refined interface. From these findings, we will identify larger usability problems or discrepancies within the information flow.

Discussion & Conclusions

This was the first major project we have worked on at SI that involved implementation. Accordingly, many of the lessons we learned are related to the development process:

  • Version Control: Towards the end of the project, we were all working on the same copy of the demo. Using a system like Subversion would have prevented us from inadvertently overwriting each other’s work.
  • MVC software architecture: Our code was messy – our view (HTML, CSS, Javascript) was mixed in the same files as our controller (PHP). This made difficult to develop and troubleshoot each part independently.
  • Comprehensive wireframes: While our lo-fi prototype was extensive, it was not complete. As a result, we did not all share the same understanding of how different sections were supposed to work. We should have completely finalized the layout and application logic before development started.
  • Challenges of user-center design: We all played dual roles as designers and developers in this project. This introduced a conflict of interest which compromised our ability to follow the user-centered design process: our design decisions were strongly influenced by what we felt were were capable of coding by the end of the semester [5].

Future Work

Our future work will focus on features, security, evaluations, and sustainability.  We hope to advance the outfit feature by adding tags, as well as the ability to edit outfits that have already been created.  Most importantly, from a social application standpoint, we will integrate the Outfit Creator and Hot-or-Not game so that people can get quantitative feedback about their choices.  Further, we hope to enhance system effectiveness by recommending other items that go with the item currently being viewed, based on popular outfits created with that item.  We also hope to add a privacy feature by allowing users to mark an item as “private.”  This will not prohibit other users from viewing an item, but it will not publicly associate the user with the marked item.

For the interface, our goal is to show images and outfits for top search results, the ability for users to upload their own pictures, and explain why a particular item is being recommended to a user. This last point is similar to the “Related to Items You’ve Viewed” and similar section headers on Amazon.com.  In order to accomplish this, we will need either a better understanding of Directed Edge, or end up having to scrap it for our own recommendation algorithm.

Other future directions include virtual directories, such that a user’s profile will show up as www.suitup.com/username. Also, Suit Up will be able to  check for duplicate item URLs when saving an item. This will minimize duplicate item clutter on our server. Adding cookies will make it so users don’t have to log in each time they visit the site. Finally, we would like to fully implement the browser plug-in or bookmarklet that allows users to more conveniently add items to their Suit Up account.

To secure the privacy of our users, Suit Up plans on encrypting passwords and having sessions expire.  We also need to re-evaluate our site’s robustness through thorough usability and user testing. It also needs much stronger error handling. We also want to ensure quality content and have developed but not implemented an item and user “flagging” system to control misuse.

Lastly, with regards to sustainability, we plan on implementing a reputation system and incorporating ads into our layout. Though not finalized, our initial thoughts with respect to a reputation system would involve feedback on users’ outfits and items, providing users with a higher status for more valuable input, similar to the SlashDot method.  A user’s activity within the system would also be taken into account, using item and outfit creation to help calculate status.

Bibliography

[1] Quarterly Retail E-Commerce Sales: 4th Quarter 2009, U.S. Census Bureau News, U.S. Census Bureau, 2010.

[2] dmreinke, Recommender Systems and why Fashion Social Shopping hasn’t worked, StyleHop.com, January 27,2009.

[3] Engeström, Jyri, Why some social network services work and others don’t — Or: the case for object-centered sociality, Zengestrom.com, April 13, 2005.

[4] Michael Rappa, Business Models on the Web, digitalenterprise.org, January, 17, 2010.

[5] John S. Rhodes, Interaction Design: The Guru Speaks, WebWord.com, March 16, 1999.