Understanding why a model makes a certain prediction can be as crucial as the prediction's accuracy in many applications. However, the highest accuracy for large modern datasets is often achieved by complex models that even experts struggle to interpret, such as ensemble or deep learning models, creating a tension between accuracy and interpretability. In response, various methods have recently been proposed to help users interpret the predictions of complex models, but it is often unclear how these methods are related and when one method is preferable over another. To address this problem, we present a unified framework for interpreting predictions, SHAP (SHapley Additive exPlanations). SHAP assigns each feature an importance value for a particular prediction. Its novel components include: (1) the identification of a new class of additive feature importance measures, and (2) theoretical results showing there is a unique solution in this class with a set of desirable properties. The new class unifies six existing methods, notable because several recent methods in the class lack the proposed desirable properties. Based on insights from this unification, we present new methods that show improved computational performance and/or better consistency with human intuition than previous approaches.

/pdf/a-unified-approach-to-interpreting-model-predictions-414asxggew.pdf

A unified approach to interpreting model predictions

A Course in Game Theory presents the main ideas of game theory at a level suitable for graduate students and advanced undergraduates, emphasizing the theory's foundations and interpretations of its basic concepts. The authors provide precise definitions and full proofs of results, sacrificing generalities and limiting the scope of the material in order to do so. The text is organized in four parts: strategic games, extensive games with perfect information, extensive games with imperfect information, and coalitional games. It includes over 100 exercises.

A Course in Game Theory

This paper provides a framework for addressing the question of when transactions should be carried out within a firm and when through the market. Following Grossman and Hart, we identify a firm with the assets that its owners control. We argue that the crucial difference for party 1 between owning a firm (integration) and contracting for a service from another party 2 who owns this firm (nonintegration) is that, under integration, party 1 can selectively fire the workers of the firm (including party 2), whereas under nonintegration he can "fire" (i.e., stop dealing with) only the entire firm: the combination of party 2, the workers, and the firm's assets. We use this idea to study how changes in ownership affect the incentives of employees as well as those of owner-managers. Our framework is broad enough to encompass more general control structures than simple ownership: for example, partnerships and worker and consumer cooperatives all emerge as special cases.

/pdf/property-rights-and-the-nature-of-the-firm-2iye4ykluw.pdf

Property Rights and the Nature of the Firm

LetN be a finite set andz be a real-valued function defined on the set of subsets ofN that satisfies z(S)+z(T)źz(SźT)+z(SźT) for allS, T inN. Such a function is called submodular. We consider the problem maxSźN{a(S):|S|≤K,z(S) submodular}.

Several hard combinatorial optimization problems can be posed in this framework. For example, the problem of finding a maximum weight independent set in a matroid, when the elements of the matroid are colored and the elements of the independent set can have no more thanK colors, is in this class. The uncapacitated location problem is a special case of this matroid optimization problem.

We analyze greedy and local improvement heuristics and a linear programming relaxation for this problem. Our results are worst case bounds on the quality of the approximations. For example, whenz(S) is nondecreasing andz(0) = 0, we show that a "greedy" heuristic always produces a solution whose value is at least 1 ź[(K ź 1)/K]K times the optimal value. This bound can be achieved for eachK and has a limiting value of (e ź 1)/e, where e is the base of the natural logarithm.

An analysis of approximations for maximizing submodular set functions--I

This important text and reference for researchers and students in machine learning, game theory, statistics and information theory offers a comprehensive treatment of the problem of predicting individual sequences. Unlike standard statistical approaches to forecasting, prediction of individual sequences does not impose any probabilistic assumption on the data-generating mechanism. Yet, prediction algorithms can be constructed that work well for all possible sequences, in the sense that their performance is always nearly as good as the best forecasting strategy in a given reference class. The central theme is the model of prediction using expert advice, a general framework within which many related problems can be cast and discussed. Repeated game playing, adaptive data compression, sequential investment in the stock market, sequential pattern analysis, and several other problems are viewed as instances of the experts' framework and analyzed from a common nonstochastic standpoint that often reveals new and intriguing connections.

/pdf/prediction-learning-and-games-31z1lqjpec.pdf

Prediction, learning, and games

An examination of a number of elementary properties of a value for the essential case. This value is deduced from a set of three axioms, having simple intuitive interpretations.

/pdf/a-value-for-n-person-games-58w5vcviy3.pdf

A Value for n-person Games

(2013). College Admissions and the Stability of Marriage. The American Mathematical Monthly: Vol. 120, No. 5, pp. 386-391.

/pdf/college-admissions-and-the-stability-of-marriage-43tjwjm34q.pdf

College Admissions and the Stability of Marriage

In a stochastic game the play proceeds by steps from position to position, according to transition probabilities controlled jointly by the two players. We shall assume a finite number, N , of positions, and finite numbers m K , n K of choices at each position; nevertheless, the game may not be bounded in length. If, when at position k , the players choose their i th and j th alternatives, respectively, then with probability s i j k > 0 the game stops, while with probability p i j k l the game moves to position l . Define s = min k , i , j s i j k . Since s is positive, the game ends with probability 1 after a finite number of steps, because, for any number t , the probability that it has not stopped after t steps is not more than (1 − s ) t . 

Payments accumulate throughout the course of play: the first player takes a i j k from the second whenever the pair i , j is chosen at position k. If we define the bound M: M = max k , i , j | a i j k | , then we see that the expected total gain or loss is bounded by M + ( 1 − s ) M + ( 1 − s ) 2 M + … = M / s . (1)

The process therefore depends on N 2 + N matrices P K l = ( p i j k l | i = 1 , 2 , … , m K ; j = 1 , 2 , … , n K ) A K = ( a i j k | …

https://www.pnas.org/content/pnas/39/10/1095.full.pdf

Lloyd S. Shapley

Papers

A Value for n-person Games

College Admissions and the Stability of Marriage

Stochastic Games

17. A Value for n-Person Games

REGULAR ARTICLEPotential Games