How to improve the label placement in scatter plot

The following builds on tcaswell’s answer.

Networkx layout methods such as nx.spring_layout rescale the positions so that they all fit in a unit square (by default). Even the position of the fixed data_nodes are rescaled. So, to apply the pos to the original scatter_data, an unshifting and unscaling must be performed.

Note also that nx.spring_layout has a k parameter which controls the optimal distance between nodes. As k increases, so does the distance of the annotations from the data points.

import numpy as np
import matplotlib.pyplot as plt
import networkx as nx
np.random.seed(2016)

N = 20
scatter_data = np.random.rand(N, 3)*10


def repel_labels(ax, x, y, labels, k=0.01):
    G = nx.DiGraph()
    data_nodes = []
    init_pos = {}
    for xi, yi, label in zip(x, y, labels):
        data_str="data_{0}".format(label)
        G.add_node(data_str)
        G.add_node(label)
        G.add_edge(label, data_str)
        data_nodes.append(data_str)
        init_pos[data_str] = (xi, yi)
        init_pos[label] = (xi, yi)

    pos = nx.spring_layout(G, pos=init_pos, fixed=data_nodes, k=k)

    # undo spring_layout's rescaling
    pos_after = np.vstack([pos[d] for d in data_nodes])
    pos_before = np.vstack([init_pos[d] for d in data_nodes])
    scale, shift_x = np.polyfit(pos_after[:,0], pos_before[:,0], 1)
    scale, shift_y = np.polyfit(pos_after[:,1], pos_before[:,1], 1)
    shift = np.array([shift_x, shift_y])
    for key, val in pos.items():
        pos[key] = (val*scale) + shift

    for label, data_str in G.edges():
        ax.annotate(label,
                    xy=pos[data_str], xycoords="data",
                    xytext=pos[label], textcoords="data",
                    arrowprops=dict(arrowstyle="->",
                                    shrinkA=0, shrinkB=0,
                                    connectionstyle="arc3", 
                                    color="red"), )
    # expand limits
    all_pos = np.vstack(pos.values())
    x_span, y_span = np.ptp(all_pos, axis=0)
    mins = np.min(all_pos-x_span*0.15, 0)
    maxs = np.max(all_pos+y_span*0.15, 0)
    ax.set_xlim([mins[0], maxs[0]])
    ax.set_ylim([mins[1], maxs[1]])

fig, ax = plt.subplots()
ax.scatter(scatter_data[:, 0], scatter_data[:, 1],
           c=scatter_data[:, 2], s=scatter_data[:, 2] * 150)
labels = ['ano_{}'.format(i) for i in range(N)]
repel_labels(ax, scatter_data[:, 0], scatter_data[:, 1], labels, k=0.008)

plt.show()

with k=0.011 yields

and with k=0.008 yields