Hulu machine learning questions and Answers series | The seventh bomb: unsupervised Learning algorithm and evaluation

I hear that Hulu machine learning is better than a winter weekend.

You can click "Machine Learning" in the menu bar to review all the previous installments of this series and comment on your thoughts and comments.

At the same time, in order to make everyone better understand Hulu, the menu "about Hulu" also made the corresponding adjustment, curious babies, brand turn up bar!

Today's content is

"Unsupervised learning Algorithms and evaluation"

Scenario Description

People have a strong inductive thinking ability, good at a lot of debris from the facts or data to find universal law, and get logical conclusions. Take the user to watch the behavior of the video as an example, there can be a variety of intuitive induction, such as from the perspective of viewing content, there are like to watch cartoons, like to watch the idol play, like to see sci-fi films, etc. from the point of view of the use of equipment, like to watch on the desktop, have a favorite mobile phone or tablet There are also like to watch on the TV and other large screen, from the use of habit, there are like to watch in the evening, have like to watch at noon, there are users watching every day, there are only on the weekend to watch users, and so on. Effective grouping of all users is important for understanding the user and recommending the appropriate content to the user. Usually this kind of problem does not have the label of the observation data or the grouping information, needs the algorithm model to seek the data intrinsic structure and the pattern.

Problem

Taking the clustering algorithm as an example, assuming that there is no external tag data, how to distinguish the two unsupervised learning (clustering) algorithm of the pros and cons?

Background knowledge: Unsupervised learning, common clustering algorithms

Solutions and Analysis

The example in the scene description is a typical clustering problem, and it can be seen that the clustering of data depends on the definition of requirements, and also relies on the method of feature measurement and data similarity of categorical data. Compared with supervised learning, unsupervised learning usually does not have the correct answer, the design of the algorithm model directly affects the final output and performance, it is necessary to find the optimal parameters of the model through multiple iterative methods. Therefore, to understand the characteristics of common data clusters and the characteristics of common clustering algorithms, it is very helpful to seek the methods to evaluate the performance of different clustering algorithms.

Features of common data clusters:

• Data families defined in the center: Such data sets tend to be spherical, and the center is usually defined as the centroid, which is a bit of the average in this data cluster. The distance from the data in the collection to the center is closer than the distance to the other cluster centers;

• Data clusters defined in density: Such data sets are significantly different densities, or dense or sparse, from the surrounding data clusters. Density-based cluster definitions are often used when data clusters are irregular or coiled together and have noise and outlier points;

• Data families defined by connectivity: there is a connection between data points and data points in such data sets, and the entire data cluster is represented as a graph structure, which is valid for irregular shapes or entangled data clusters;

• A concept-defined data cluster: All data points in such data sets have some common properties.

Features of the common Clustering algorithm:

• Partitioning clustering: Data objects are divided into non-overlapping data families, where each data point is in a single data cluster;

• Hierarchical clustering: Data families can have sub-clusters, and data families with multiple (nested) subgroups can be represented as tree structures;

• Fuzzy clustering: Each data point belongs to a data cluster with the membership weight of 0~1;

• Full/incomplete Clustering: whether to assign a data family to all data points.

Because of the diversity of data and requirements, no algorithm can accommodate all data types, clusters, and applications, and it seems that each situation may require a different assessment metric. For example, K-means clustering usually needs to be evaluated with SSE (Sum of Square Error), but density-based data clusters may not have to be spherical and SSE will fail completely. In many cases, judging the result of clustering algorithm depends strongly on subjective interpretation. However, the evaluation of clustering algorithm is necessary, it is one of the important parts of clustering analysis.

The evaluation of the merits and demerits of clustering algorithms can usually be summarized into the following five aspects of analysis:

1. The ability to identify the existence of non-random cluster structures in the data;

2. Ability to identify the correct data clusters in the data;

3. Ability to evaluate data being properly clustered;

4. Ability to identify the pros and cons of two data clusters;

5. Assess the difference between an objective data set and an estimate;

Assuming that there is support for external callout data, the 5th will be transformed into a supervised learning problem, and the cluster structure found by the clustering algorithm is directly measured to match the structure of the labeled data. Assuming there is no external callout data, based on the 1~4 points listed above, you can test the clustering of clustering algorithms for different types of data clusters, as shown in the following:

Figure 1. Whether the observed error is monotonically changing with the increase of the number of cluster classes

Figure 2. Effect of observation error on clustering results

Figure 3. Observing clustering accuracy of neighbor data cluster

Figure 4. Observing the performance of clustering algorithms when dealing with large data density differences

Figure 5. Observe the clustering accuracy when dealing with different size data types

Scaling issues

In the above answer, we introduced five kinds of evaluation of two clustering algorithm performance merits and demerits, then what are the common indicators can be used to calculate and identify the advantages and disadvantages of clustering algorithm? This paper presents several possible data cluster patterns, and defines the evaluation index to show the ability of the interviewer to solve and analyze the problem. In fact, there are a number of measurement indicators, the following list of three in the data compact or data cluster can be separated degree of measurement, more indicators please refer to the literature [1], specifically described as follows:

• RMS standard deviation (RMSSTD), which measures the homogeneity of the cluster:

• R-Square (r-square) to measure cluster variance:

• Improved hubertγ statistics that assess cluster differences through inconsistencies in data pairs:

This includes:

Next Topic Preview

"Intensive Learning"

Scenario Description

Game is one of the most representative and suitable application areas of reinforcement learning, which covers almost all the elements of reinforcement learning, such as the environment: the state of the game itself, action: User operation, Robot: program, Feedback: score, win and lose, etc. Playing video games by entering raw pixels is one of the hallmarks of AI maturity. Atari (Atari) is a the Seventies or eighties of the 20th century smash hit computer game, similar to the domestic red and white machine game, but the picture elements to be more simple. Its simulator is relatively mature and simple, it is very appropriate to use the Atari game to test reinforcement learning. The application scenario can be described as: on the discrete timeline, each moment you can get the current game screen, choose to send a command to the game console (up and down, fire, etc.), and then get a feedback (reward). There is no way to use traditional methods directly due to the large state space of the reinforcement learning based on the primitive pixels. Thus, in 2013 DeepMind proposed the deep reinforcement learning model, and began the combination of deep learning and intensive learning [2].

Traditional intensive learning mainly uses q-learning, while deep reinforcement learning also uses q-learning as the basic framework, which changes the q-learning's corresponding steps to the depth form and introduces some techniques, such as experience replay (experience Replay) to accelerate convergence and improve generalization capabilities.

Problem description

What is deep reinforcement learning, and how does it differ from traditional reinforcement learning, and how to use it to play video games?

Reference documents:

[1] Liu Y C, Li Z M, Xiong H, et al understanding of internal clustering validation measures. In:proceedings of IEEE ICDM ' ten, Sydney, 2010. 911-916.

[2] Antonoglou, I., Graves, A., Kavukcuoglu, K., Mnih, V., Riedmiller, M.A., Silver, D., & Wierstra, D. (2013). Playing Atari with deep reinforcement learning. CoRR, abs/1312.5602.

Hulu machine learning questions and Answers series | The seventh bomb: unsupervised Learning algorithm and evaluation