![](https://cdn.myportfolio.com/e2b3badb-d40c-4696-a4c0-f11ba2141dcf/e32eb75d-636e-43f1-a552-1a824df3a2b3_rw_1920.png?h=1d4c75e89f8372823c7ed601addc2804)
PROBLEM
User Recommendations
With the rise in technology utilizing ML and AI, recommendations have started to drive aspects of user centric preferences with systems. One of the fundamental goals of AI/ML recommendations is to drive higher user engagement with applications and ecosystems. The accuracy of these recommendations depends on the quality of past user data that is collected from their interactions with the system. This collection of user data defines the user preferences for generating accurate recommendations that are relevant to the user. This relevancy being targeted for higher user engagement, leads to more transactions for the service provider and better preferences for the user.
The problem with this data driven approach is the assumption that all user interactions are correlated to user preferences. Some user interactions might be unrelated to their preferences and these interactions can create outlier data thereby decreasing the accuracy of recommendations.
User Feedback
In current recommendation systems, one of the ways of correcting inaccurate recommendations from unrelated outlier data is with User Feedback. The user can explicitly give feedback as to how accurate their recommendation is. Users can rate a recommendation on a scale or by hitting a like or dislike button. This feedback will become an additional data point that will be used for future recommendations.
The disadvantage of this approach is that the outlier data set is still part of the user data and can continue to affect the accuracy the future recommendations irrespective of what user feedback is given.
![](https://cdn.myportfolio.com/e2b3badb-d40c-4696-a4c0-f11ba2141dcf/8c6b2b3c-ed73-479c-97c8-97578dbd4d29_rw_1200.png?h=3672512454a83c97646a29c4deb7be05)
Problem Statement
Current methods of improving accuracy of User Recommendations with User Feedback are not efficient. They do not directly exclude the outlier data for future recommendations and continued existence of this data can cause inaccurate recommendations to be generated in future.
SOLUTION
The proposal is a framework called ‘Transparency of Recommended Experiences’ (TRUE), that combines User Data Transparency, Explainable AI and User Feedback to improve the accuracy of User Recommendations. TRUE ensures unrelated historic user data is not used for recommendations by allowing users to manipulate and remove unrelated historic data points that could generate inaccurate results. This is achieved by exposing the user to their historic data that was used to generate a particular recommendation.
TRUE brings User Data Transparency by exposing Users to their own contextual User Data that was used for generating Recommendations (Fig. 2A and 2B) and allowing an understanding of what data generated their Recommendations. TRUE uses Explainable AI by presenting this contextual user data in the form of a visual timeline (Fig. 2B), so that Users can get a meaningful perspective of their data. TRUE also provides Users with tools to filter out data points based on events such as locations, user actions, search terms, sites visited, media viewed, etc.
If Users gets inaccurate Recommendations, TRUE utilizes User Feedback to let them examine and remove unrelated data interactions from the system and ensure it is not used for future Recommendations (Fig. 2C and 2D). This rectification of User Data will guarantee the accuracy of future Recommendations.
![](https://cdn.myportfolio.com/e2b3badb-d40c-4696-a4c0-f11ba2141dcf/69932ad9-98ea-4186-8236-8010d18e8c0d_rw_1920.png?h=8ef18f067b6983fadccddf0dd3668fcc)
HOW TRUE WORKS
Current recommender systems apply their own complex algorithms with user data for generating curations. The recommendation process generally consists of layers (Fig 3) that are involved between the user activity and the recommendation. These layers are:
1. User Activity: This layer consists of all the User’s actions and conducts that are the source, leading to recommendations. These activities could include location details, content consumed, interactions with content, content shared, voice recorded, or any aspect of the user’s daily timeline that contributed to a particular recommendation. These User activities maybe stored on local devices or on the cloud.
![](https://cdn.myportfolio.com/e2b3badb-d40c-4696-a4c0-f11ba2141dcf/63bb062e-25fc-4895-ad72-7a012d7fd473_rw_1200.png?h=27398bb0b809470096dffd32d49ed0d4)
2. Recommender System: This is the intermediate layer between the user activity and the recommendation. The user activities tracked by the system are categorized as data sets using the complex algorithms implemented by the respective recommender system. Different combinations of these data sets would be used to generate relevant recommendations. The algorithms and technology used in this layer would be respective to the service provider giving recommendations i.e. YouTube, Netflix, amazon, etc.
3. Recommendation Layer: This final layer consists of all the recommendations and suggestions displayed to the User via a User Interface. This is the only layer exposed to the user, where they will see the Recommendation.
![](https://cdn.myportfolio.com/e2b3badb-d40c-4696-a4c0-f11ba2141dcf/eb723488-fb35-4838-9798-3de83fc43606_rw_1920.png?h=e7727bbe5157eac0bae162c0e4184286)
TRUE’s fundamental goal is to connect the end user with the Recommendation Layer and User Activity. This would allow the User to view historic data that was used for the Recommendation in a manner that the User can understand, and then make modifications to it that would suit their preferences directly. This produces more relevant Recommendations that leads to higher user engagement and more transactions for the service provider.
TRUE achieves this by introducing two new layers (Fig 4) into the existing Recommendation Process. These segregated layers give users an explanative link between the recommendation and the activity that contributed to that recommendation and then take action if needed. The layers consist of the following:
1. TRUE User Control: This layer works concurrently with the existing Recommendation Layer. This layer is part of the User Interface where the User can interact with TRUE to exposes contextual historic data used for Recommendations. The dotted lines in Fig 4 and Fig 5A shows how the TRUE framework will interact with the TRUE Firewall and then the Recommender Systems to fetch this User data. The TRUE User Control then presents this complex data in a ‘Timeline’ view where User Activities are sorted based on date and time, along with the type of Activity (search terms, locations, websites viewed etc.). A ‘Timeline’ view makes this complex recommendation data easily consumable for the User and it is complemented with tools to filter by Activity Types if they want to look for specific type of information. The User can then delete events that they feel are not correlated to their preferences.
This ability to see User Data in a meaningful manner and then modify it forms the crux of TRUE. The Users can also understand what all data points are tracked about them, and if they are not comfortable with a certain type of tracking, then the TRUE User Control has the capability of directing the User to the phone’s privacy settings and make changes if needed.
2. TRUE Firewall: This layer behaves like a hub for the TRUE framework that connects the Recommender Systems and User Activities with the TRUE User Control (Fig 5A). If required, this layer can incorporate a real firewall that regulate the tracking and recording of User Activities by Recommender Systems in accordance with Local Privacy Laws like GDPR. This layer has two main roles to play:
A. The core system that TRUE User Control interacts with directly for modifying and deciding what user data should be used by the Recommender systems. When the User requests to see the data used to generate a Recommendation, the TRUE User Control will request the TRUE Firewall to expose the contextual User Activities related to that Recommendation. If the User deletes a particular activity, then the Firewall will not allow that activity to be used in future Recommendations by the Recommender Systems.
B. To act as an intermediary between the Recommender System and the User Activity. It can ensure that the Recommender Systems will only get data that is allowed by the User allows and by the legal privacy frameworks such as GDPR or equivalent local privacy laws that are applicable. The User will have a choice within the TRUE framework (in the UI, Fig 5B) to apply or not apply the local privacy law as rules for TRUE to enforce for data sharing with the Recommender Systems.
The TRUE Firewall layer could use Blockchain technology to ensure extra safety and anonymity of User Data.
![](https://cdn.myportfolio.com/e2b3badb-d40c-4696-a4c0-f11ba2141dcf/89da65f2-2478-4459-878a-5f648ec2c16a_rw_1200.png?h=fd53e88a9084bd19a9ecd9ae8fb668f9)
![](https://cdn.myportfolio.com/e2b3badb-d40c-4696-a4c0-f11ba2141dcf/957f80ff-f34d-4648-8142-5420f72f69b2_rw_1920.png?h=d0521d30c26d6e8dbad9525265a3434b)
FEATURES
1. User Data Transparency
Current Examples - Popular systems that have User Data driven Recommendations do not reveal any actual User Data that was used for generating those Recommendations. This has lead to a lack of transparency between Recommender Systems and Users and understanding how their data is tracked.
- Google does not provide any necessary information (Fig 6A) as to why an ad is shown to the User other than a list of possible reasons like age or interests and not the actual data points.
- Amazon also provides does not provide meaningful information but an unclear line stating ‘Based on your recent activity we thought you might be interested in this’ (Fig 6B). There is no relevant user activity shown to the User.
- Netflix tries to bring a closer relation between the Recommendation and the User Activity (Fig 6C) but with very basic understanding – ‘Because you watched ABC, you will like XYZ’.
- Facebook provides an understanding of what the Advertisers are looking for (Fig 6D). It creates a correlation between the Recommended Ad and User by providing what the Advertiser was looking for and match it the User Data that was used.
![](https://cdn.myportfolio.com/e2b3badb-d40c-4696-a4c0-f11ba2141dcf/352af9c6-07b1-4653-821a-4be01204089c_rw_1920.png?h=3375746c9bd3fbbf6af1d1afc54f74ed)
User Data Transparency with TRUE
As seen in the current examples, there is a lack of clarity and transparency for the User as to how their data is used to create these Recommendations. The TRUE framework intends to bring User Data Transparency by directly revealing the complete User data used to generate a Recommendation. Every Recommendation can have an indicator along with it in the UI (the ‘T’ icon next to the Ad information in Fig 7A) that opens the TRUE Framework disclosing the User Data. This complex data is shown in the UI (Fig 7B) as a ‘Timeline’ where User ‘Activities’ are sorted based on date and time, along with the type of Activity (search terms, locations, websites viewed etc.).
Exposing this information brings User Data Transparency and helps Users to understand how their data is tracked and consumed by Recommender Systems.
![](https://cdn.myportfolio.com/e2b3badb-d40c-4696-a4c0-f11ba2141dcf/66a6910f-f56f-46e0-ac4b-864592ff36bd_rw_1920.png?h=9413397fa23235b296201430c5ff48f9)
2. Explainable AI
Current Examples - Recommender Systems generally do not provide detailed explanations as to why a particular recommendation is given to user. If they do, it is mostly a generic explanation or a redirect to the User’s interest or Profile page.
- Google redirects an ad explanation to the User’s profile personalization page (Fig 8A) and not provide any contextual information related to the Recommendation.
- Facebook and Instagram does not provide any meaningful or contextual information about why a Recommendation (ad) was shown to the user. It provides a lengthy generic explanation (Fig 8B & 8C) detailing how they track and generate ads.
- Netflix tries to bring a closer relation between the Recommendation and the User Activity (Fig 8D) but with very basic understanding – ‘Because you watched ABC, you will like XYZ’.
![](https://cdn.myportfolio.com/e2b3badb-d40c-4696-a4c0-f11ba2141dcf/46a1c2d6-471f-43c5-ba60-ea23b2933c59_rw_1920.png?h=5ffec84299bee268a57cb1646b33146a)
Explainable AI with TRUE
The first interactive ability TRUE provides to the user is an explanation of the recommendation into its respective user activities. The TRUE framework presents contextual historic User Data in the UI (Fig 9) as a ‘Timeline’ where User ‘Activities’ are sorted based on date and time, along with the type of Activity (search terms, locations, websites viewed etc.). A ‘Timeline’ view makes this complex recommendation data easily consumable for the User and it is complemented with tools to filter by Activity Types if they want to look for specific type of information. Each ‘Activity’ shows when the activity occurred, and briefly describes how and what happened.
TRUE will keep the explanation of Recommendations independent of the technologies and algorithms that were used to generate it. It will only analyze the User Data used and then present it to the User in the most meaningful manner. TRUE’s goal is to present and explain as to what User data was used for a Recommendation and not how the data was converted to a Recommendation.
![](https://cdn.myportfolio.com/e2b3badb-d40c-4696-a4c0-f11ba2141dcf/3c8825b1-8362-4a5e-ab17-34c124334418_rw_1200.png?h=e155fb8f543606a86fc3a9eea9e47475)
2. Explainable AI
Current Examples - Almost all Recommender Systems get User Feedback through a questionnaire with a generic set of questions and answers to the User and this answer becomes a new data point for the Recommender System. The outlier User Activity that triggered the inaccurate Recommendation still exists in the system and could affect future results as well.
- Google triggers a questionnaire related to regulatory issues (Fig 10A).
- Facebook and Instagram asks the User why they don’t want to see an ad with a predefined set of answers (Fig 10B & 10C).
- Netflix asks the User to like or dislike a feature (Fig 10D) that they have watched and use that as a data point for future Recommendations. Even if the Users dislikes it, the User Activity of watching that feature still exists in its system and could be used for future Recommendations.
![](https://cdn.myportfolio.com/e2b3badb-d40c-4696-a4c0-f11ba2141dcf/b6ba3dcf-d707-4e11-9f39-8267a2ce7d2f_rw_1920.png?h=b46d756a9c2e4187da4068f15b779c88)
User Feedback with TRUE
TRUE gives the interactive ability to the User to fine tune their data that is consumed for generating the recommendation. Some recommendations may be triggered by unsolicited user activity, and it would be of help if the user is explicitly involved in making an amendment. The User can then delete events that they feel are not correlated to their preferences. With TRUE, the user can backtrack such unwanted suggestions right up to the user activity that generated it. After investigating the set of activities that triggered an undesirable recommendation, the User can choose to:
A. Deleting activity from system: This feature allows the user to remove a particular activity from all the recommender system data sets consuming it. There are various unintended activities that are monitored by the system that in turn generate undesired recommendations. These could be the map location sensed, speech recognition tag words recorded, etc. Users can examine and remove unrelated these data interactions from the system and ensure it is not used for future Recommendations (Fig. 11A). This rectification of User Data will guarantee the accuracy of future Recommendations.
![](https://cdn.myportfolio.com/e2b3badb-d40c-4696-a4c0-f11ba2141dcf/631b9326-45e7-4958-a3b5-7c5c70c7cd23_rw_1920.png?h=fc6b5aeee3c8f7f9f3dd5b970e69ec29)
B. Deleting activity from context: This feature allows the user to remove the undesired activity from the particular recommendation being investigated. This control is provided if the user wishes that a particular activity should not be contributing to a respective recommendation (Fig 11A). However, the same activity will still be consumed by other data sets that generate recommendations.
C. Change System Privacy Settings: TRUE allows Users to connect User Activities that’s are tracked by Recommender Systems and their device Privacy Settings. If a User feels a particular type of User Activity should not be tracked by the Recommender System, then TRUE provides a shortcut directly to device Privacy Settings and make changes that they want.