Posts tagged AI

New Research by MIT Uncovers the Behavioral and Network Features for “Follow Back” on Twitter

 

Interesting insights and exploration from Dylan Walsh, who tells the tale of Taumid Zaman, MIT professor, who tries to tries to get a follow from Taylor Swift and ends up with a new tool for information warfare. To influence someone on social media, first you need them to follow you. New research uncovers the behavioral and network features that make that happen.

Reference: http://mitsloan.mit.edu/newsroom/articles/solving-twitters-follow-back-problem/?utm_source=mitsloantwitter&utm_medium=social&utm_campaign=followback

It was 2014. Taylor Swift had recently released her single “Shake It Off.” She was now a certifiable pop star and Tauhid Zaman, associate professor of operations at MIT Sloan, wondered if he could get her to follow him on Twitter. Swift had about 60 million followers; he had fewer than 1,000. She represented a global empire; he was an academic. A long shot, yes, but these odds were precisely what motivated the question. “I wanted to know what makes people follow you back,” Zaman said. “Celebrities have a wall around them, but their weaknesses on social media are the people they follow.”

Could he somehow use a celebrity’s friends on Twitter — Swift’s hair stylist or sound engineer — to open the gates to her inner circle? He dubbed this the “follow-back problem,” and he solved it with his students at MIT. The first step of this process was to understand the underlying dynamics of follows on Twitter. For instance, what kinds of Twitter interactions matter the most when trying to get followers? And do overlapping social networks actually help build connections? If they do, then to what degree do they help?

Read More

Zaman tested these questions using a group of Twitter bots posing as artists. Each bot was designed to promote a real artist’s work through Twitter’s three main interactions: following, retweeting, and replying. By tracking these interactions and the responses, Zaman was able to methodically probe and quantify the behavior of users on Twitter.

Two basic principles emerged: first, intuitively, those who don’t follow many other people are unlikely to follow you back, while those who follow a lot of people are likely to follow you if you follow and retweet them. Second, social overlap matters. If Swift follows somebody who, in turn, follows Zaman, then Zaman has a greater chance that Swift will follow him. This boost follows a predictable pattern where the friend of my friend is my friend.

But simply understanding these relationships wasn’t Zaman’s goal.

“We’re engineers, and so we wanted to design a system around this insight,” he said.

By the time he and his team got to work on this, though, “Shake it Off” had become much less interesting than the world’s most famous Twitter user, President Donald Trump. What, he wondered, would be the most promising path to get a follow from @realDonaldTrump?

Zaman ran a model to find the optimal sequence of interactions to garner a follow from Trump, assuming you could only interact with 10 or 20 of his connections. (As the number of interactions gets larger, Zaman said, a Twitter account becomes increasingly suspect, looking more like a bot than a real person.)

He found that targeting the right people in the right order made a follow from the president four to five times likelier than a random approach; and if the follow-back campaign expanded to include friends of friends, then the likelihood jumped even higher. In the end, by targeting a network of 200 individuals on Twitter associated with Trump and the people he follows, Zaman found that he could increase the chance that the president would follow him back by an order of magnitude compared to an uncoordinated campaign. The chance was still small, about 2 percent in his calculations, but it still showed the impact of following people in a smart way.

What does this have to do with democracy and counterterrorism? 
As frivolous as this result may seem, Zaman’s work is both timely and relevant to core questions of democracy and counterterrorism, and more generally information warfare. Consider the involvement of Russian bots on Twitter and Facebook now understood as a concerted effort to sway results of the 2016 election.

Or consider the social media accounts created by organizations like the Islamic State group, which has very effectively expanded membership through these channels. Given this social media landscape, cracking the follow-back problem is the first, essential step for infiltrating an adversary’s network. By targeting certain Twitter accounts, for instance, Zaman believes it may be possible to spread information that dampens the effect of foreign actors in domestic elections, or that counters the recruitment propaganda spread by IS.

This prospect, he admits, is equal parts exciting and scary. While there is plenty of good that can come out of these tools — getting people to exercise, eat their vegetables, stop joining IS — there is an obvious dark side.

“In my opinion, this can be far more dangerous than conventional weapons which have a fixed blast radius,” Zaman said.

While social media tools don’t present direct physical threats, they can powerfully influence the opinions of a whole country; they can, in Zaman’s analogy, have a tremendous blast radius.

“These are weapons, and I’m building efficient ways to use the weapons, so this has to be handled with care,” he said.

Zaman hasn’t yet used the modeling results from this work to pursue a Twitter follow from Swift and Trump, but he is considering giving it a try. And as for the follow-back problem, he is planning on incorporating it into a full-fledged social network counter-measure for influence campaigns by hostile state and non-state actors.

Or, as he puts it, he is “developing the tools for the next generation of information warfare.”

This is the first in a three-part series examining new work about Twitter, influence, and bots by MIT Sloan associate professor Tauhid Zaman.

 

 

.

8 Fundamentals for Achieving AI Success in the Supply Chain

Sharing point of view from  Greg Brady is the CEO and founder of One Network Enterprises, a global provider of a secure and scalable multi-party business network. For more information, contact the author at gbrady@onenetwork.com or visit http://www.onenetwork.com. Enjoy!

Source: 8 Fundamentals for Achieving AI Success in the Supply Chain – Supply Chain Management Review

There’s a lot of buzz and hype about artificial intelligence (AI) in supply chain management (SCM). That’s understandable given its potential. AI can offer a huge benefit to supply chain managers, but only if it is based on solid fundamentals that take into account the diverse and dynamic nature of today’s modern supply chains. More importantly, it needs to consider the availability of the timely and accurate data needed to make smart decisions.

Before addressing what AI can do, it is critical to first understand what it is. In the simplest terms, AI is intelligence exhibited by machines, or when machines mimic or can replace intelligent human behavior, such as problem solving or learning. In essence, AI is machines making decisions whether that is deciding which chess piece to move where, or how to adjust an order forecast based on changing demand.

Despite its benefits, when looked at through the lens of a supply chain executive, AI is relatively useless unless it’s able to add value to support better decision-making.

Why AI Hasn’t Delivered in SCM

In the race to use AI, many companies have made attempts to implement it, but the results have been disappointing. This is because typical SCM systems today:

  • Require armies of expensive planners
  • Run complex engines at each step in the process and at each node in the supply network
  • Are usually in conflict with other functions and/or partners
  • Miss huge opportunities hidden in the network because they are locally sub optimized
  • Work on stale data and thus promote bad decisions
  • Use dumbed-down, over-simplified problem models that do not relate to the real world

These SCM limitations have severely suppressed return on AI investments. For example, typical Retail/CPG supply chains still carry 60-75 days of inventory. The average service level in the store is about 96 percent, with promoted item service levels much lower at the 80 percent range. The Casual Dining segment on the other hand, carries around 12 – 15 days of inventory with relatively high waste and high cost-of-goods-sold. So, unless AI can make a significant impact on these metrics, it’s simply not delivering.

Key Requirements for AI in Supply Chain Management

Having worked with hundreds of supply chain executives, on dozens of software implementations, I’ve studied the AI issue a lot. What I have found is there are eight criteria that are required for a successful AI implementation. Miss one of these and you’ll be lucky to achieve mediocre outcomes, but when you meet them all, you can indeed achieve world class results. For the AI solution to offer optimal value in supply chain, it important to ensure the following:

1. Access to Real-Time Data 
To improve on traditional enterprise systems with older batch planning systems, new AI systems must eliminate the stale data problem. Most supply chains today attempt to execute plans using data that is days old, but this results in poor decision-making that sub-optimizes the supply chain, or requires manual user intervention to address. Without real-time information, an AI tool is just making bad decisions faster.

2. Access to Community (Multi-Party) Data 
The ability to access data outside of the enterprise or, more importantly, receive permission to see the data that is relevant to your trading community, must be made available to any type of AI, Deep Learning or Machine Learning algorithms.

Unless the AI tool can see the forward-most demand and downstream supply, and all relevant constraints and capacities in the supply chain, the results will be no better than that of a traditional planning system. Unfortunately, this lack of visibility and access to real-time, community data is the norm in over 99 percent of all supply chains. Needless to say, this must change for an AI tool to be successful.

3. Support for Network-Wide Objective Functions
The objective function, or primary goal, of the AI engine must be consumer service level at lowest possible cost. This is because the end-consumer is the only consumer of true finished goods products. If we ignore this fact, trading partners will not get the full value that comes from optimizing service levels and cost to serve, which is obviously important as increased consumer sell-through drives value for everyone.
A further enrichment of the decision algorithm should support enterprise level cross-customer allocation to address product scarcity issues and individual enterprise business policies. Thus, AI solutions must support global consumer-driven objectives even when faced with constraints within the supply chain.

4. Decision Process Must Be Incremental and Consider the Cost of Change
Re-planning and changing execution plans across a networked community in real time can create nervousness in the community. Constant change without weighing the cost of the change creates more costs than savings and reduces the ability to effectively execute. An AI tool must consider trade-offs in terms of cost of change against incremental benefits when making decisions.

5. Decision Process Must Be Continuous, Self-Learning and Self-Monitoring
Data in a multi-party, real-time network is always changing. Variability and latency is a recurring problem, and execution efficiency varies constantly. The AI system must be looking at the problem continuously, not just periodically, and should learn as it goes on how to best set its own policies to fine tune its abilities. Part of the learning process is to measure the effectiveness “analytics,” then apply what it has learned.

6. AI Engines Must Be Autonomous Decision-Making Engines
Significant value can only be achieved if the algorithm can not only make intelligent decisions but can also execute them. Furthermore, they need to execute not just within the enterprise but where appropriate, across trading partners. This requires your AI system and the underlying execution system to support multi-party execution workflows.

7. AI Engines Must Be Highly Scalable
For the supply chain to be optimized across an entire networked community of consumers to suppliers, the system must be able to process huge volumes of data very quickly. Large community supply chains can have millions if not hundreds of millions of stocking locations. AI solutions must be able to make smart decisions, fast, and on a massive scale.

8. Must Have a Way for Users to Engage with the System
AI should not operate in a “black box.” The UI must give users visibility to decision criteria, propagation impact, and enable them to understand issues that the AI system cannot solve. The users, regardless of type, must to be able to monitor and provide additional input to override AI decisions when necessary. However, the AI system must drive the system itself and only engage the user on an exception basis, or allow the user to add new information the AI may not know at the request of the user.

AI in the Real World Today

Sounds good in theory, but how does it work out in practice? Now that we have addressed the key fundamentals, let’s look at how some actual companies have achieved applying these criteria.

For instance, one of the major problems in Casual Dining is anticipating and meeting demand for the restaurants, corporate owned or franchised. This is especially important during Limited Time Offers (LTOs). Using the eight criteria outlined above, a global, casual dining company connected to a real-time, multi-party network, and was able to rapidly achieve their objective function – excellent customer service at the lowest cost.

The company constantly monitors Point-of-Sale (POS) data, and is using AI agents to recognize and predict consumption patterns of consumers. In addition, intelligent AI agents create the demand forecast and then compare it to the actual demand in real-time. When there is significant deviation, the agents make the decision to adjust the forecast, and additional agents adjust replenishments. They then propagate those adjustments across the supply chain to trading partners in real time at all times considering the cost of change and the propagation impact.

This drove a remarkable improvement in forecast accuracy. During promotions, the company achieved over 85 percent forecast accuracy at the store level and even higher at the DC level. This represents at least a 25 percent improvement over traditional approaches.

Intelligent agents also optimize restaurant orders autonomously by recognizing the impact of projected restaurant traffic trends and impact on LTOs and therefore the orders. The system runs on an exception basis but allows the managers to review the decision criteria and override orders where the managers may have local information such as inventory issues or local store traffic issues. This has resulted in much faster order placement and order accuracy of over 82 percent, which reduces both inventory and waste dramatically while increasing service levels to the consumer. This is a significant improvement to all other known implementations in the marketplace.

Because the algorithms are highly scalable, they are processing over 15 million stocking locations continuously throughout the day.

Prior to the AI-based, multi-party execution system, restaurant managers had to interact with nine different ordering systems and manually create their own orders based on general guidelines, rules of thumb, and spreadsheet-based or manual calculations.

With AI implemented on a sound foundation, this company can now anticipate, manage, and serve demand at the lowest possible cost. During LTO’s, when demand fluctuations would overwhelm a restaurant manager, intelligent agents monitor demand in real time, and autonomously orchestrate the supply chain to align supply with demand. Thus, the company can meet its goal and maintain high service levels while reducing cost to serve.

These are not isolated results. Also in the food marketplace, another CPG-Retail implementation achieved 99 percent in-stock, in-store, with 25 days of supply (DOS) across the supply chain.  The inventory results are less than half the standard DOS in this marketplace and 3 percent points higher in in-store in-stocks
AI-based solutions are being deployed at two large automotive tier one suppliers with results ranging from 16 – 40 percent reductions in inventory as well as significant reductions in expedited freight costs.

AI Delivers Value in SCM Today

As you can see, laying the proper groundwork for AI pays huge dividends. There’s no doubt that AI offers even greater promise in the future, but, as these results show, there are significant benefits and dramatic results waiting for companies that focus on the fundamentals and put AI to use today.

The beauty of AI-based solutions is that they learn and drive continuous improvement over time. They get more precise and sophisticated as they gather more data and more experience. The sooner you start, the better the results you’ll see in future, and the further ahead you will be. With the right AI solution in place, you can outpace your competitors today, and be well positioned for reaping even bigger rewards of AI’s promise tomorrow. ~Greg Brady

[/toggle]