On an average, we make about 70 conscious decisions in a typical day. Our brain makes dozens of micro decisions (boolean decisions) every single second. Choices, sub-choices, super-sub-choices, tiny little micro decisions building up decision trees and forests to arrive at a macro decision. It takes seconds sometimes and sometimes even days to years in arriving at one. We cannot get by our daily lives without deciding on something. Decision making is an integral part of our life, some even consider it an art and others call it a proficiency. Whether it is an art or a proficiency, one thing is certain, there is always someone at the center of it, the decision maker. Whether you believe in the simulation hypothesis or in the law of attraction, It’s all about the person at the center, the decision maker, a human, an agent or an automated system (an entity), who picks a particular choice from the available options.
The Decision Maker & The Best Choice
The decision maker or the entity is deemed “intelligent” or “smart” if there’s an underlying criterion or rationale behind picking a particular choice. The degree of intelligence or smartness is determined by the ability to pick the “best” choice from the available options (obviously, the entity or the selection is not considered intelligent if the choice is random). Now, what does this underlying criteria or rationale function on? IMO, there are three core aspects that are absolutely essential in picking the best choice, first, ‘available information’, which helps in identifying, eliminating and narrowing the options (sometimes, the noise from the observations increases the effort to eliminate options, which results in poor quality of decisions). Second, the ‘time’, the total available time to make the decision which provides a sense of urgency or latency to the act. Finally, the ‘reward’, which mostly is the primary motivation for making the decision. For the purpose of this article, let’s focus on the first part of this rationale, the ‘available information’.
The Seeker, Available Information & The Giver
How do we gather the required information before picking the best option? It depends on the situation, isn’t it? if it’s a repetitive situation then, we depend on a previously built mental model, in other words ‘experience’. If it’s a new situation then we build newer models with available information and depend on the experience model for future occurrences. We have a tendency to depend on our personal preferences for simpler choices but, when it comes to complex decisions, we tend to gather more information. One of the quick hacks our brain employs in solving a complex decision problem is depending on an expert’s recommendation. Why not? We are social beings, why not use social information (especially when we face a lack of necessary information sitch) to solve a troubling decision problem. But wait, how do we filter these social recommendations? This is where “trust” comes into the picture. Trust, a dependable model our brain builds over a period of time, based on past experiences to hack a complex decision-making task. Now, for certain situations, the entire process of decision making boils down to the binary relationship between ‘the seeker’ (the decision maker) and ‘the giver’ (the expert who gives the recommendation).
Core Attributes Of An Expert
What are the core attributes of an “expert” who can give us the ‘best’ recommendation? Being “expert” at the subject is certainly an important criteria, we have also established “trust” to be a qualifying factor in our mental model, “accessibility” is another attribute that can directly appeal to the time component we have talked about earlier and yes, should be “intelligent” to understand the seeker and the “context” to give the most “relevant” and “accurate” recommendation to qualify it as the “best” recommendation from the available options.
What Happens In The Real-World?
Let’s understand this model from a trade and business context. When you enter a store, a real store (offline store with walls, furniture, merchandise, and salespeople), salespeople act or approach as the experts in a store to guide us in our decision-making process, that is their role, they often receive training on the sales process, they have knowledge on business profit numbers and they are trained to respond to our queries about the product and to help us make the best choice. From a business perspective, salespeople are not scalable, nor economically viable so, businesses invented a non-technical hack called POGs (‘Plan-o-grams’ or ‘Planograms’) to arrange items to increase visibility on high-profit and high-frequency items to increase revenues and optimize the store spaces. Some online stores borrowed the same idea to display their high-profit, high-frequency items on their homepage layouts to optimize their sales. The problem with online stores is the absence of a definitive sales guide and population of their online stores with the abundance of options. Sometimes, there are too many options and too many choices available and users do not like to invest their time in going through the entire catalog to find the best choice. An obvious answer to this problem is an intelligent recommendation system, a system that can mimic the role of a salesperson, a system that can reduce the workload on users who are overwhelmed by the number of available options. Not just help in reducing options; typically, in a customer journey, customers go through stages like need, awareness, research, comparison, decision & purchase. Other than the initial need and final purchase stage, an intelligent recommendation system should be able to assist customers through the stages of awareness, research, and decision.
Every Successful Product Or Business Has A Strong Recommendation Engine At Its Core.
Amazon’s — “Customers who bought this item also bought…”.
Netflix’s — “Other Movies You May Enjoy…”
Spotify’s — “Recommended songs…”
Google’s — “Visually Similar Images…”
YouTube’s — “Recommended Videos…”
Facebook’s — “People You May Know…”
LinkedIn’s — “Jobs You May Be Interested In…”
Coursera’s — “Recommended courses…” and
Waze’s — “Best Route…”
are all results of strong recommendation systems at the core of these businesses. The impact of these recommendation systems is immense from a business standpoint as well.
According to this McKinsey’s article, 35% of Amazon’s revenue and 75% of what people watch on Netflix comes from their recommendation engines.
Macro Components Of An Intelligent Recommendation Engine
Now, from the context of building a product or a service with a strong recommendation system in place; it may be easier to build a fast, automatic, heuristic based system but, how can we put the decision maker, the user in control? How can we build an intelligent recommendation ‘engine’ that can mimic the dynamic interactive human nature between the seeker and the giver? How can we build an engine to understand the context, capture the immediate feedback on the recommended items and also calculates the long-term benefits for you, like a friend? Before that, let’s understand one of the important macro-components of an intelligent recommendation system — filtering models. What is a model? According to this “Models Will Run The World” article by Steven A. Cohen and Matthew W. Granade:
“A model is a decision framework in which the logic is derived by algorithm from data, rather than explicitly programmed by a developer or implicitly conveyed via a person’s intuition. The output is a prediction on which a decision can be made. Once created, a model can learn from its successes and failures with speed and sophistication that humans usually cannot match.”
In short, models are the true source of power behind robust business systems and recommendation (filtering) models are one of the most popular research subjects for data scientists today. There are 5 types of recommendation filtering models that I’ve come across: (the word ‘filtering’ is used because, filtering is a fundamental common requirement when dealing with large sets of data, filtering helps in focussing on relevant data).
Types Of Filtering Models
1. Popularity Filtering Model: This is the most basic and simplest of all models, the recommendations are based on the number of views, likes, ratings or purchases. Most popular items on the platform (product) are recommended to all the users. The limitation of this model is that it is uni-dimensional, i.e., the recommendations stay the same for every user.
2. Collaborative Filtering Model: This is a user-user collaborative model, it is more like an extension to the popularity model but, micro and more personalized. Information is collected on the users to find similar users and the recommendations flow based on the interactions of one user on the platform to another similar user. The limitation of this model is that the performance of the system slows down as the user base grows.Sometimes users are clustered into pairs and groups to speed things up. A trade-off product owners have to make with user personalization part in the favor of product speed.
3. Content Filtering Model:The idea here is similar to the collaborative model except that the correlation is found between items and not users. That is the reason why this model is also known as the item-item collaborative model. However, it is extremely crucial for product owners to define a strong similarity framework. The advantage of this model is that the relationship between items can be more stable (depends on the similarity framework) than users and in general items are a smaller base compared to user base so, the speed of the system stays good even if the user base grows.
4. Association Rule Mining Model: The idea here is an extension of the content filtering model. The difference is that the focus is shifted from the “items” to a “session”. What it means is, the items that appear together in the same ‘session’ are recommended to other users across the platform. For example, an association rule could be like: 80% of users who like item A and item B also like item C; and 20% of all users like all three items. Another rule could be like: 70% of all items liked by user X and user Y are also liked by user Z and 30% of all items are liked by all three users. The concept is based on association rules here, and the correlation is formed based on the ‘appearance’ in the same session. The limitation of this model is that it is less personalized compared to collaborative or even content filtering models.
5. Hybrid Filtering Model: As the name suggests, this model combines two or more filtering techniques to achieve better performance. Usually, collaborative filtering methods are combined with content filtering methods in layered, weighted, switched, mixed and cascaded approaches to achieve desired results. The advantage of this model is that it is more powerful, more stable and more relevant compared to the other models.
How Can We Improvise Hybrid Filtering Model With Machine Learning?
So far, the above models have been performing pretty well in the product and business world. However, the fundamental assumption around the concept of these models needs a relook. For example, if we take the robust hybrid filtering model, the core assumption here is that — the proposed ‘best’ recommendation to a user is the one that other users with a comparable profile in a comparable state who chose the same or a similar product. In other words, the assumption is that statistically, the recommended product is what the user would’ve chosen with or without the recommendation. The system is picking the best. But, would the user really would’ve chosen the same product without the recommendation? Do recommendations change user behavior or do they merely shortens the path of an inevitable choice? Essentially, the subject of recommendation reduces to statistical analysis of understanding users, products and their relationship. The goal of the system design should be to reduce the information or data that is useless and irrelevant for effective decision making and to stimulate buying action by the user. The question is, how can we improvise? how can we design a system that can learn to understand the nuances of user behavior, take various aspects of business into consideration and then statistically analyze the best recommendation, with the right balance?
Implicit & Explicit Data Capture At Product Design Level
Since the subject is now boiled down to statistical analysis and relationships and when we talk about modern statistical analysis and techniques to identify patterns from large data sets, machine learning and data mining techniques become an obvious choice. Data is the most essential component of any model and machine learning models thrive on data. However, handling data can be quite complicated, and there are lots of things that can go wrong like, inconclusive data, noise in the data, the readiness of data, bias in data collection and confidence intervals. Data is also extremely crucial to keep the ‘entropy’ of the entire system in check. The ‘purer’ the data-sets and their sub-sets the minimal the entropy of the system will be in the long run.
To generate personalized, specific and tailored recommendations it is essential to understand the users accurately and the first step in understanding them accurately is to collect the right data about them. User data can be collected ‘implicitly’, which is to record user behavior and recognize patterns and ‘explicitly’, which is to collect data by specific inquiry. Below is an example of a mobile app that I have designed. On the left side screen, you will see the implicit approach, where calendar, contacts and location data is captured to understand the user implicitly and identify early patterns. On the right side of the screen, you will see the explicit interrogative approach where information on user’s ‘interests’ is captured in dynamic bubbles moving up from the bottom of the screen, newer bubbles moving up are based on user’s earlier selection (interactive idea here is to gamify a mundane task of interrogation and make it fun for the user to engage and share more information). The goal is to learn from the user reaction about the relevancy of the items to the user and to form a relationship with the content that will follow in the next screens.
Once we have capture basic requisite data and form data-sets, we should be able to generate the first layer of recommendations (Reco.1) see from Fig.6 below and also possibly identify “similar users” (if we have defined framework in place). The implicit inquiry that started with the onboarding screens can flow into the main content screens, implicit and explicit inquiries can be designed well into the sessions and record user behavior to identify patterns (if any) for future actions. Based on these captures we can form various data-sets like data-set#4 (see from Fig.6 below), this, in turn, forms the input to generate the second layer of recommendations (Reco.2) like a loop.
An example of an explicit inquiry can be seen in the screens below, capture and real-time analysis followed by relevant content population. While data set#4 is generated with user’s inputs, data set#5 here is system generated, following an algorithm in identifying similar items and populating content accordingly as the third layer of recommendations (Reco.3).
While Reco.1 is unidimensional, Reco.2 and Reco.3 are multidimensional and based on system learning, learning from similar users and similar items. The goal of the system design here is to increase the effectiveness of data by making it more relevant to stimulate buying action by the user. If the product has special offers as a feature, the complexity of the task even goes a few notches up with frequent varying prices added into the equation (like the screen on the right in the image below).
A good recommendation engine should be in a position to learn, adapt and deliver the best recommendation always. This is possible if we can model the engine to analyze the historical data about the user with respect to the item or items the user may be browsing at a given moment using real-time analysis. Real-time analysis can infuse a sense of ‘proactiveness’ to the system that is different from the present ‘reactive’ nature of recommendation engines that are predominantly depended on analysis of just historical data. One of the quick examples for real-time analysis is Google’s Penguin 4.0. Google updated their site ranking algorithm, Penguin 4.0 to real-time (year 2016). What this means is, Penguin’s data is refreshed in real-time, so the changes in page ranking appear much faster than before (year 2016).
A natural conversational interface that allows simple interactions would be extremely valuable in capturing real-time data. The screens below that I’ve designed for this purpose demonstrates exactly that, an evolution of interface design, a concept of a virtual agent backed by an intelligent recommendation engine in the background capable of even sensing user’s emotional state to adapt and suggest the best recommendation that fits right into user’s need, sometimes even before the need arises. The power of prediction goes a long way in demonstrating product value and building trust. It’s more like an intriguing thought experiment. But, surely we’re getting closer every day to such dream states and dreamy product experiences.
Inclusion Of Systemic Smarts
So the learning of the recommendation engine happens in a step-by-step manner. Goal oriented learning from historical data, learning from real-time interactions, learning from trial and errors in the environment, with rewards and penalties to achieve better accuracies is precisely the task of Reinforcement Learning (RL), one of the most fascinating disciplines of machine learning. Reinforcement learning acts as a multidimensional learning multiplier over this model. According to one of the most standard and comprehensive works on RL, “Reinforcement Learning: An introduction”, by Richard S. Sutton and Andrew G. Barto there are two main elements and three key sub-elements for an RL model. Two elements — an agentand its environment. Three sub-elements, a policy, a reward signal, and a value function.
“Elements: An ‘agent’ is the active decision-making agent and its ‘environment’ is the space or structure within which the agent seeks to achieve a goal despite uncertainty. Sub-elements: A ‘policy’ defines the learning agent’s way of behaving at a given time. A ‘reward signal’ defines the goal in a reinforcement learning problem. A ‘value function’ specifies what is good in the long run.”
Pertaining to our model and context, systemic active decision-making agent (not to be confused with the human decision maker we discussed earlier) and its environment here is the recommendation engine model (from Fig.6), extending from data set#1 to Reco.3. A policy is a mapping of each state and the action. For instance, action to throw Reco.2 as suggestions to a new user based on the state of interactions from a similar old user. The final reward of the model here is the purchase of the product by the user. A more immediate reward is the reaction of the user, like sharing the product with a friend or liking it or bookmarking it or adding to favorites. Value function could be the accumulation of various rewards starting from the first interaction with the system to extrapolating and correcting over the future. The ability to course correct the value function to an incremental value over time can enable the system to add both values to the business and delight to the user in the long run.
Once the engine or system reaches beyond satisfactory levels of accuracy, the next step could be to appropriate the dissimilarity quotient of the items. This can help the system to incorporate the surprise element into recommendations. IMO, getting the surprise element right is finding the holy grail of a recommendation engine.
The Holy Grail
The surprise element in this context is the ability of the recommendation engine to recommend a dissimilar item from the known items to a given user in a given situation. The more dissimilar and non-obvious an item is from the known, the more surprise the user can be. The surprise element as a component fits into the content filtering section of the hybrid model because arriving at a dissimilarity quotient is the other side of understanding similar items. The serendipitous nature of discovering an interesting item which user might not have otherwise discovered will boost the delight and help the system to form a deeper bond with the user. This will also help build the ‘trust’ element we spoke of earlier in this post (trust is a fascinating topic to study under game theory). If we have to define ‘serendipity’ as a property of an intelligent recommendation engine, the two most fundamental factors that serendipity is depended is ‘relevancy’ and ‘timing’. From here, we can go deeper into evolving equations for surprise element, serendipity, trust and how it all fits into the overall equation of the gargantuan recommendation engine. Turning a surprise recommendation into a serendipitous occurrence in the context of the recommendation engine is equivalent of the hand of God move (game#4, move#78) in the Go match between AlphaGo & Lee Sedol. Life could be a box of chocolates when an intelligent recommendation engine is handing the box.